PHD Thesis

The Determinants of Total Factor Productivity:
The High-Performing Asian Economies Revisited
Arlene Locsin Garces Ozanne
A thesis submitted for the degree of

Doctor of Philosophy
of the University of Otago, Dunedin,
New Zealand
July 2001
ii
iii
~ ABSTRACT ~
Despite the recognised importance of total factor productivity (TFP) in
explaining the remarkable economic performance of the High-Performing Asian
economies (HPAEs), the study of TFP and its determinants has received little
attention in the empirical literature on economic growth. This thesis attempts to make
a contribution in this area.
We hypothesise that the degree of openness of an economy, the different roles

of government, and human capital are the major factors that influence TFP. We also
carefully examine the different ways of incorporating human capital into our analysis.
Different proxies for the role of government are used to capture the different effects of
government on the economy. We also acknowledge the importance of using better
proxies for the degree of openness of the economy. We use different types of data
(cross-sectional, time-series and pooled time-series and cross-sectional data) and
appropriate econometric tools to test our hypotheses.
Results from our cross-sectional analysis show that government intervention,

measured in terms of government consumption expenditure, price controls and
government ownership of enterprises, is a statistically significant influence on the
TFP levels of countries. Contrary to our hypothesis, government intervention in
terms of total government consumption expenditures is a positive influence on TFP.
The other forms of government intervention, e.g., ownership and control of
enterprises and price controls, are, as hypothesised, negative influences on TFP, but
these results are quite fragile depending on the other variables included the regression
equation.
There is some evidence that the degree of openness of the economy
influences TFP levels. A less open economy, measured in terms of either the black
market exchange rate premium or taxes on international trade as a percentage of the
total trade sector, is a negative influence on TFP levels. The effect human capital is
not statistically significant.
Our time-series analysis of the individual HPAEs,
however, shows that only government intervention, measured in terms of noncomplementary expenditure, is a statistically significant influence on TFP, and that
only for Malaysia.
Granger causality tests for the joint causal effects of the
hypothesised determinants, on the other hand, show that these variables jointly
iv
Granger cause TFP levels in four of the five HPAEs examined (i.e., Malaysia,
Singapore, South Korea and Thailand). Finally, estimation of our pooled time-series
and cross-sectional data, allowing for fixed country effects, provides support for a
long-run level relationship between TFP and our hypothesised determinants.
Although the results of this study appear to be somewhat fragile, suffering persistently
from the problem of not having sufficient time-series data, all of our estimations are
supplemented by a series of diagnostic tests that we find lacking in many of the
previous empirical studies.
~ ACKNOWLEDGMENTS ~
I would like to thank my supervisors, Professor Dorian Owen, Dr Stephen
Knowles and Dr Robert Alexander, for their encouragement, patience and valuable
advice, which helped shape and improve this thesis.
I would also like to thank all the members of the Economics Department of the
University of Otago for their support and goodwill; and to all the lecturers who
allowed me to tutor and lecture on their courses.
I am also grateful for the constructive comments I received from the participants
of the Economics Departments seminar series, the Association of Pacific Rim
Universities (APRU) 2nd Annual Doctoral Students Conference (Auckland, February
2001), the New Zealand Association of Economists (NZAE) Conference
(Christchurch, June 2001), the Econometrics Society Australasian Standing
Committee (ESASC) Inaugural Intensive Workshop for Young Scholars (Hamilton,
July 2001).
I am also grateful for the financial support from the Macandrew-Stout

Postgraduate Scholarship, the Donald-Reid Scholarship in Economics, and other
grants from the Economics Department.
I owe a special debt to my family in the Philippines, who despite not being here
with me, always prayed for and encouraged me in my pursuit of this degree. I would
also like to thank my extended family in New Zealand for their concern and
cheerfulness.
Finally, I have left until last, someone to whom I would like to extend my
special thanks, my husband, Rick. Thank you very much for your encouragement and
support. Thank you for always being there for me.
Any errors and omissions in this thesis are my sole responsibility.
vi
vii
~ TABLE OF CONTENTS ~
Abstract
iii
Acknowledgments
Table of Contents
vii
List of Tables
xiii
List of Figures
xv
Chapter
1. Introduction
2. The HPAEs Economic Growth Experience
2.1. Introduction
2.2. What is TFP?
2.2.1. FA and TFP growth
2.2.2. TFP and its determinants within a single growth equation
17
2.2.3. TFP from growth accounting
18
2.2.4. TFP as residuals from a production function
21
2.3. Government and Economic Growth
26
2.3.1. To intervene or not to intervene
27
2.3.2. Government intervention in the HPAEs
28
2.3.3. Analysing the empirical evidence
30
2.4. Openness and Economic Growth
43
2.4.1. Establishing the link between openness and economic growth
44
2.4.2. Analysing the empirical evidence
46
2.5. Human capital and growth in the HPAEs
57
2.6. Conclusions
59
viii
3. Explaining Differences in Total Factor Productivity Across Countries
61
3.1. Introduction
61
3.2. Estimating TFP levels
63
3.2.1. Hall and Jones (1999) TFP levels
63
3.2.2. An alternative measure of TFP levels
65
3.3. Correlation analyses
66
3.3.1. Data
66
3.3.2. Correlation analysis results
71
3.4. TFP and its determinants
80
3.4.1. Framework
80
3.4.2. Regression results
82
3.5. Robustness and Diagnostic tests
85
3.5.1. Robustness tests
85
3.5.2. Diagnostic tests
87
3.5.3. TFP estimations without the influential observations
93
3.6. Summary and conclusions
96
Appendices to Chapter 3
A.1
Explanation of GOE Ratings
101
A.2
Explanation of Price Ratings
102
1988 TFP levels (in logs)
103
TFP estimation results, using TFP from a production function

with human capital-augmented labour, with human capital and
interaction terms as regressors in the TFP equation
106
C.1
Diagnostic test results
107

with human capital-augmented labour, without human capital and
interaction terms as regressors in the TFP equation
108
D.1
109
TFP estimation results, using all data from Hall and Jones (1999)
For calculating TFP from a production function without human
E.1
capital-augmented labour
110
111
ix
F
TFP estimation results, using all data from Hall and Jones (1999)
For calculating TFP from a production function with human
capital-augmented labour
112
F.1
113
List of countries included in the main TFP regressions
114
4. Testing for Cointegration and Causality for the HPAEs
115
4.1. Introduction
115
4.2. Data
118
4.3. TFP levels for the HPAEs
125
4.4. The relationship between the Granger causality concept

and cointegration
133
4.4.1. The Granger causality concept
133
4.4.2. Cointegration, error correction models (ECM) and causation
134
4.5. Analysis of the time-series properties
137
4.5.1. Unit root testing procedure
137
4.5.2. Choosing the appropriate lag length
138
4.5.3. Unit root test results
139
4.6. Cointegration and error correction model (ECM) representation

for the HPAEs
146
4.6.1. Johansens multivariate cointegration and ECM representation
146
4.6.2. Testing for cointegration in the HPAEs
149
4.6.3. OLS estimation results for South Korea
152
4.7. Testing for Granger causality
152
157
H
TFP levels in the HPAEs, using the latest Collins and Bosworth
(1996) dataset, without human capital in the production function
161
H.1 Indonesia
161
H.2 Malaysia
162
H.3 Singapore
163
H.4 South Korea
164
H.5 Thailand
165
x
I
Unit root test results
166
I.1
Indonesia
166
I.2
Malaysia
168
I.3
Singapore
170
I.4
South Korea
172
I.5
Thailand
174
Plots of Human for Malaysia, South Korea and Thailand
176
Results of applying Johansens (1988) procedure

for testing cointegration
177
K.1 Indonesia
177
K.2 Malaysia
179
K.3 Singapore
181
K.4 Thailand
183
OLS estimations of the ECMs
185
L.1
Indonesia
185
L.2
Malaysia
190
L.3
Singapore
195
L.4
Thailand
198
5. The determinants of Total Factor Productivity:

Evidence from a Panel of HPAEs
203
5.1. Introduction
203
5.2. Panel data analysis
204
5.3. Panel data
211
5.4. Panel unit roots
214
5.5. Unit root test results
218
5.6. A bounds test approach to the analysis of level relationships
222
5.6.1. Pesaran et al.s (2001) testing strategy
222
5.6.2. Bounds test results
225
5.6.3. Bounds test results for the existence of level relationships

in a TFP equation with country-specific dummy variables
227
xi
5.6.4. Bounds test results for the existence of level relationships
in a TFP equation using de-trended variables
233
237
M
Complete data set for the estimations (1973-1990)
240
ADF Unit root test results for the individual HPAEs
243
A sample list of Shazam commands for estimations of pooled data

with lagged explanatory variables
246
Error correction form of the ARDL (0,0,3,0,3,0) TFP equation
247
248
6. Summary, conclusions and suggestions for further research
249
References
255
xii
xiii
~ LIST OF TABLES ~
Table
2.1
Selected growth accounting studies on East Asia
11
Selected empirical work on government and economic growth
31
Selected empirical work on openness and economic growth
48
Government and Openness Indicators
68
Spearmans rank correlation coefficients for TFP, output,

government and openness measures
72
Spearmans rank correlation coefficients for the alternative TFP measures
78
Spearmans rank correlation coefficients for government and

openness variables
80
TFP estimation results
84
89
TFP estimation results without the influential observations
94
Diagnostic test results for regressions without influential observations
95
Data for multivariate cointegration and causality analysis
126
TFP growth rates for selected HPAEs
129
Estimates of TFP growth rates from selected empirical studies
132
Summary of Unit Root Test Results
145
Granger Causality Test Results
157
Selected panel data studies on growth
208
Results of IPS Panel Unit Root Tests
220
Error Correction form of the ARDL (1, 5, 3, 3, 2, 1) TFP equation

(Dependent variable: lnTFPit, 1973-1990)
229

For the ECM form of the ARDL (1, 5, 3, 3, 2, 1) TFP equation
230

(Dependent variable: lnTFPit, 1973-1990) (variables are de-trended)
235

236
6
7
8
xiv
xv
~ LIST OF FIGURES ~
Figure
1
2
Average share of government expenditure in GDP (%)

(in 1985 international prices)
39
Level of Openness: East Asian vs. Latin American economies
45
xvi
~ CHAPTER ONE ~
Introduction
The period spanning the 1960s to the early 1990s saw the economic rise of a
group of East Asian countries: the High-Performing Asian economies (HPAEs) of Hong
Kong, Singapore, South Korea, Taiwan, Malaysia, Indonesia and Thailand.1 For over
three decades, the remarkable economic performance of these HPAEs was unparalleled
in the world. The average annual growth rate of (real) GNP per capita of the HPAEs
from 1965 to 1990 was 6.7%. Over the same period, the average annual growth rate
was only 2.7% for the OECD countries, 0.8% for the economies in Latin America, and
0.7% for the Sub-Saharan African economies (Villegas, 1998).
Despite the financial crisis that struck the East Asian economies in late 1997 and
1998, the fact remains that never before had countries been transformed so quickly from
poor, underdeveloped economies into industrialised economies.
For this reason,
analysis of the success of the HPAEs continues to attract considerable attention in the
literature on economic growth. We note that the economic growth experience of the
HPAEs, particularly of the four tiger economies (Hong Kong, Singapore, South Korea
and Taiwan), has already been extensively studied and, according to Temple (1997, p.
280), diminishing returns may be setting in. However, despite the weight of research,
the experience of the HPAEs remains open to various interpretations.
A number of different explanations have been put forth in the research on the
nature of the remarkable economic performance of the HPAEs. Some researchers claim
that the extraordinary output growth of the HPAEs was largely due to an equally
impressive rate of factor accumulation (FA), while others say that total factor
productivity (TFP) growth was the key to the HPAEs success (see Chapter 2 for
details). The World Bank (1993a), Page (1994) and Stiglitz (1996) among others, stress
the importance of having an open trading regime, educated and skilled workforce and
1
The term HPAE was first used by the World Bank (1993a) to describe these economies, plus Japan.
However, Japans economic takeoff began in the 1950s, with growth rates slowing by the 1970s. For this
reason, we will use the term HPAE to describe the remaining seven economies, which experienced high
growth rates up until 1997.
2
judicious government policies in promoting growth in East Asia, although many other
factors can be named. How these factors should actually be accounted for in empirical
growth models is another issue that needs further clarification. Empirical evidence
appears to raise more questions than answers. Diagnostic tests are often not reported
and the use of different types of data (cross-sectional, time-series or panel data) also
leads to substantially different findings.
This thesis sets out re-examine the HPAEs experience, by focusing on TFP and
its determinants, rather than on output growth per se, for reasons discussed more fully in
the succeeding paragraphs. We carefully consider the theoretical links between TFP
and three factors (government, openness and human capital) that are regarded as
important influences on it. We also aim to provide information on the robustness of (or
lack of robustness) the effects of these factors on TFP. In particular, we investigate the
sensitivity of the empirical findings to the type of data used in the analysis, proxies used
to capture the conceptual equivalents of the hypothesised determinants of TFP, and the
econometric techniques used.
A review of growth accounting studies that look into the issue of how much
output growth in the HPAEs is due to TFP growth and how much is due to FA (see
Chapter 2) clearly shows two opposing views. One group, typified by Young (1995),
argues that East Asias growth performance is largely due to massive factor
accumulation (FA), while the other group (e.g., Kawai, 1994; King and Levine, 1994;
Sarel, 1997, and Drysdale and Huang, 1997) points out that, after accounting for FA,
something else, i.e., TFP growth, is as important for explaining much of the growth in
output of the East Asian economies. Although the importance of TFP in the remarkable
economic performance of the East Asian countries has been recognised in these studies,
identifying the determinants of TFP has received little attention in the empirical
literature. For this reason, this thesis attempts to make a contribution in this area.
More generally, productivity growth has long been argued to be the ultimate
determinant of long-run growth (see for instance Arrow, 1962; Uzawa, 1965; Romer,
1986; Lucas, 1988, and the Solow-Swan neoclassical growth model (Solow, 1956 and
3
Swan, 1956)). Furthermore, it has also been argued that differences in levels and
growth rates of TFP can help explain a major part of differences in levels and growth
rates of output per worker across countries (Klenow and Rodriguez-Clare, 1997;
Prescott, 1998; Hall and Jones, 1999; and Easterly and Levine, 2000). Hence, in order
to obtain a better understanding of the HPAEs remarkable economic performance in
the long-run, we need to look beyond FA, and focus on TFP.
Attempts at understanding the determinants of TFP, however, have to deal with

definitional, theoretical and empirical modelling issues. For instance, what does TFP
actually mean? What does it measure? As Lipsey and Carlaw (2001, p. 3) point out,
TFP clearly means different things to different informed observers. We present more
fully the different interpretations of TFP in the next chapter. For now, we simply
note that for the purposes of this study, we subscribe to the conventional definition of
TFP, i.e. that TFP growth measures improvements in technology and efficiency.
In this study, we must also deal with the issue regarding the apparent lack of a
theoretical framework within which to analyse TFP and its determinants (Prescott,
1998), which may be partly a consequence of not having a solid definition of TFP in the
first place. Because of this, it is far from obvious which variables actually need to be
included in our analysis. Durlauf and Quah (1999) point out that in most empirical
studies, the choice of variables is often ad hoc. In fact, their critical survey of the
empirical growth literature shows that well over 90 potential growth determinants
have been examined. In addition, there is likely to be significant heterogeneity in the
relevant parameters across countries.
Because of this, Brock and Durlauf (2001)
suggest that a useful exercise may be to first identify a relatively homogeneous group of
countries. Using data from these countries, one can then draw inferences that more
reasonably apply to all members of the group. The grouping of countries will also have
implications for the choice of variables to be included in the analysis.
Throughout this study, our analysis is directed at examining the levels of TFP
instead of TFP growth rates because levels capture differences in the long-run
relationships among the variables of interest (e.g. TFP and its determinants). Once the
4
variables are differenced, valuable long-run information about the between-country
level relationships is lost; hence, it is more appropriate to study levels rather than
growth rates.
As mentioned previously, the remarkable economic performance of the HPAEs,

for over three decades from the 1960s, set the HPAEs apart from the rest of the world.
Based on the recent literature on the economic growth of the HPAEs (see Chapter 2),
we find that government, the degree of openness of the economy and human capital
have often featured in these growth studies. In most of these studies, either output
growth or output growth per capita is regressed on these variables. However, in this
study, we regress TFP (levels), rather than growth of output or output per capita, on
these variables. This is because, as Miller and Upadhyay (2000) point out, simply
treating possible determinants of output growth (e.g., trade openness) as factor inputs
may be conceptually inaccurate because the included determinants may only have
indirect effects on output. It may be that the possible determinants of output growth
affect the efficiency of the traditional factors of production (labour and physical capital),
i.e., they affect the level of TFP. For this reason, it may be preferable to treat these
variables (e.g., trade openness) as determinants of TFP rather than of output.
Although other possible determinants of TFP have been suggested in the

theoretical literature, the decision to focus on three (government intervention, openness
and human capital) is based primarily on the empirical literature on the economic
growth of the HPAEs that suggests that these are among the major factors that play a
crucial role in these countries growth experience (see for example, World Bank,
1993a). We present next the theoretical links between economic growth and implicitly
TFP since the implications for TFP are similar, and these three major variables.
The market failure argument may perhaps be the most commonly quoted
justification for government intervention, but in no way does government intervention
guarantee economic success. Nonetheless, armed with appropriate policies, government
can drive the economy not only towards an efficient allocation of resources, but more
importantly, toward sustained economic growth. Governments can try to affect the rate
5
of growth of the economy through various policies directed at improving resource
allocation and overall factor productivity (e.g. through tax concessions for firms in the
more efficient sectors of the economy); encouraging the flow of investment, both
domestic and foreign, into the country (e.g. through careful management of interest
rates and foreign exchange rates); promoting technical progress through public and
private investments in research and development (e.g. by guaranteeing patents,
royalties, copyrights, etc.); and promoting the development of human capital (e.g.
through educational subsidies, setting of educational standards to ensure quality of
education provided in state and private educational institutions and better access to these
institutions).
However, the extent to which government interventions have been
successful in increasing long-run economic growth rates has been very different across
countries.
Thus, although the theoretical links between government intervention and

economic growth have been widely studied in the experience of the HPAEs, much
disagreement as to how exactly governments have contributed to the success of these
economies still remain. Chapter 2, section 2.3, examines a number of theoretical and
empirical studies on government intervention and economic growth.
Concerning openness and economic growth, the standard argument in support of

the hypothesis that openness positively affects economic growth rests on the notion that
openness or trade leads to improvements in TFP, as a result of either static or dynamic
gains from trade. Static gains from trade arise as a result of reallocating resources to the
production of goods for which the economy has a comparative advantage.
By
specialising in the production of goods that the economy can produce at lower
opportunity cost than other countries, the economys consumption possibilities expand.
That is, the country can consume outside its production possibilities frontier (PPF).
Dynamic gains, on the other hand, occur when trade improves the production
possibilities of an economy, i.e. an outward shift of the PPF, because of improvements
in technology. A more open economy can also gain from economies of scale since the
international market is much larger than the domestic economy.
6
In addition to static and dynamic gains, a country can also benefit from
competitive gains from trade. In order for an economy to maintain its position in the
international market, its export industries are pressured to keep production costs low,
operate in a more efficient manner, and maintain high quality export products, and/or
continuously improve products (Emery, 1967). This competitive pressure from the
international markets also stimulates the flow of market innovations, managerial skills,
etc. to the exporting country, which are likely to induce TFP growth. These competitive
pressures force the economy to operate on the PPF rather than inside it. Similarly,
competitive pressures also keep import competing (domestic) industries on their toes.
Chapter 2, section 2.4, examines further the links between growth and openness in a
number of theoretical and empirical papers on this issue.
Lastly, the main contribution of human capital development (generally and often
crudely measured in terms of school enrolment ratios) to the growth process is to
increase the level of cognitive skills possessed by the labour force, and as a
consequence, to improve the marginal productivity of the workforce (Benavot, 1989 see
also Chapter 2, section 2.2). Based on this, many economists assume that the HPAEs
initial advantage in education facilitated their rapid growth. Data from Deinger and
Squire (1997) show that primary and secondary enrolment ratios in the HPAEs on
average (94% and 30% respectively), are much higher compared to the average in other
Asian and Latin American countries.
In this thesis, we aim to provide some support for the importance of the effects
of this selected set of variables (government, openness and human capital) on the TFP
levels of the HPAEs. Recognising that empirical studies often yield different results
depending on the type of data used, we conduct our analyses using cross-sectional,
time-series and pooled cross-sectional and time-series data. We do this in order to
investigate the implications of using such different types of data in a broadly common
framework. The use of either type of data has its own advantages and disadvantages,
and no one type can be conclusively regarded as being better than the others.
7
In the search for the determinants of growth, current empirical practice has
tended to focus on the use of either cross-country data (e.g. Kormendi and Meguire,
1985; Barro, 1991; Pack and Page, 1994, and Young, 1994, etc.) or time-series data
(e.g. Jung and Marshall 1985; Chow, 1987; Serletis, 1992, and Kocherlakota and Yi,
1996, etc.)
More recently, panel data have also been used in economic growth
regression analyses (e.g. Knight et al. 1993; Barro and Lee, 1994; Taylor, 1995; Islam,
1995; Lee et al. 1996; Thomas and Wang, 1996; Kneller et al. 1999, and Bleaney et al.
2001, etc.). Some authors have also attempted to compare results from the use of
different types of data, i.e. cross-sectional vs. time-series data as in Ram (1986),
Easterly and Rebelo (1993) and Edwards (1992), or cross-sectional vs. panel data as in
Harrison (1996). Results using different types of data appear to yield different results.
For instance, Harrison (1996) finds that only one out of the seven openness variables
she used positively affects growth in her cross-sectional regressions, while six out of the
seven openness measures affect growth in her panel data regressions.
Cross-sectional analyses have become an important starting point for discussions

about economic growth (see for example Romer, 1986 and Lucas, 1988). Within the
context of TFP studies, one aim of running cross-country regressions is to use the
variation in TFP across different units (i.e., countries) to understand its determinants.
One advantage of this approach is that cross-sectional regressions can produce more
accurate long-run coefficients while by contrast, time-series variation presents only
shorter-run responses (Baltagi and Griffin, 1983 and Temple, 1999a). However, there
are a number of problems associated with cross-sectional regressions, beginning with
the most frequently cited objection about the assumption that different countries
(regardless of how very different they are, e.g. Sierra Leone and Singapore) are drawn
from a common surface, i.e., the problem of parameter heterogeneity.
In this case, the use of time-series analyses may offer a solution to this
objection. Focusing on a single country over time may provide more information about
the determinants of growth or TFP than estimating some average relationship over a
diverse group of countries. Time series analyses also avoid problems associated with
converting some variables into a common currency and compiling statistics across a
8
wide range of countries in a consistent fashion. Problems such as omitted variable bias
and simultaneity bias are also better dealt with within a time-series context than in
cross-sectional studies. However, time-series studies are not without problems. First,
time series studies are limited by the availability of historical data. Many important
variables such as population growth and school enrolment data are often interpolated
from census data, which are usually available only in five-year intervals. When the time
period is short (e.g. less than thirty years for some variables (Temple, 1999a)), it is often
difficult to ascertain long-run relationships between variables.
The effects of the
business cycle may also influence relationships determined. While it is possible to

prevent short-run business cycle effects from affecting the results by including long lags
of the independent variables, this may lead to a problem of not having enough degrees
of freedom left for the estimations.
The relatively recent (i.e., over the last decade) availability of comparable data
for a wide range of countries over several time periods e.g., Summers and Hestons
(1991) Penn World Tables and consequently, the use of panel data techniques has
enabled researchers to overcome many of the problems associated with either crosssectional or time-series studies. Although problems such as those mentioned previously
are not likely to disappear in a panel data context, panel data studies nonetheless, may
suffer to a lesser degree because both cross-sectional and time-series variations are
considered in the estimations. Temple (1999a) notes two main advantages of using
panel techniques over pure cross-sectional or time-series techniques in the study of
growth. These are, first, panel studies allow one to control for omitted variables over
time, and second, several lags of regressors can be used as instruments where required,
thus reducing measurement error and endogeneity biases. A more detailed discussion of
the advantages of panel data analysis is presented in Chapter 5.
Although the use of panel data in growth studies does offer several advantages,
we do not dismiss the usefulness of either cross-sectional and time-series studies. In
fact, in this thesis, we use these different approaches to test the robustness of our
findings regarding TFP and its determinants.
9
The rest of the thesis is organised as follows:
Chapter 2 examines the more significant empirical studies on the HPAEs

economic growth experience. We review growth accounting studies that attempt to
determine how much of the HPAEs growth is attributable to FA and TFP growth. We
then examine the different interpretations and ways of calculating TFP, as well as
approaches to studying its determinants. We also examine the roles of government,
trade openness and human capital, which, as mentioned previously, have often featured
in the empirical literature on East Asian economic growth as among the major factors
that play a crucial role in the HPAEs growth experience.
Chapters 3 to 5 examine the determinants of TFP using cross-sectional, timeseries and pooled cross-sectional and time-series data, respectively.
In Chapter 3, we derive a new set of estimates of TFP levels based on the

methods used by Hall and Jones (1999) and examine the correlations among TFP and
our hypothesised determinants of TFP: government, openness and human capital. After
examining the underlying pattern of simple correlations, multivariate TFP equations
with various government and openness variables and human capital as regressors are
estimated for a wide cross-section of countries. A series of diagnostic tests on each of
the equations estimated is conducted to examine the reliability and robustness of results.
We also acknowledge, as previous studies have shown (e.g. Harrison, 1996),

that the significance of variables being studied can differ substantially depending on the
particular type of dataset used (i.e., cross-section, time-series or pooled). Hence, we
also examine the links among TFP, government, openness and human capital using
time-series data for five HPAEs,2 focusing on the causal relationships among these
variables in Chapter 4. Time-series analysis also has some distinct advantages over
pure cross-section analysis, as discussed more extensively in this chapter. Hypothesised
cointegrating and causal relationships are analysed within multivariate cointegration and
causality frameworks following Johansens (1988) cointegration procedure and
2
The five HPAEs in our panel are: Indonesia, Malaysia, Singapore, South Korea and Thailand. We are
unable to include Hong Kong and Taiwan in our panel because of data constraints.
10
estimation of error correction models (ECMs). We follow a three-step procedure to test
for causality, similar to Engle and Grangers (1987) procedure.
In Chapter 5, we attempt to re-examine the results from our cross-section and

time-series studies by using panel data for five of the HPAEs and appropriate methods
to deal with non-stationarity in the panel (pooled) data.
We also apply Pesaran
et al.s (2001) recently developed bounds testing strategy to test the existence of longrun level relationships between a dependent variable and a set of regressors when it is
uncertain whether the variables are stationary or non-stationary.
Finally, the summary and conclusions are presented in Chapter 6. This study,
based on a careful review of both the theoretical and empirical literature on the
economic growth experience of the HPAEs and analysis of different types of
econometric data, offers an insight into the fragility of relationships between TFP and
some key factors that contributed to the HPAEs economic success. Of the three key
factors we examine, only the government-related variables appear to have a significant
influence on TFP regardless of the type of data used. However, although careful checks
have been made to ensure that findings in this study are robust, this study closes with an
acknowledgment of the weaknesses that remain in the current empirical practice of
analysing TFP and its determinants, but affirms that this does not imply that research
into this area is futile.
11
~ CHAPTER TWO ~
The HPAEs Economic Growth Experience
2.1
Introduction
This chapter examines the more significant empirical studies on the HPAEs
economic growth experience, with particular emphasis on the methods employed and
the strengths and weaknesses of the analyses.
As mentioned in the introductory chapter, although the economic growth

experience of the HPAEs, particularly of the four tiger economies, has been intensively
studied, the experience of the HPAEs remains open to various interpretations.
However, the different arguments and the main points raised in more recent studies can
be classified into four major debates: (1) Is economic growth in the HPAEs due mainly
to factor accumulation (FA) or to gains in total factor productivity (TFP)? (2) What
role do governments play in the phenomenal experience of the HPAEs? (3) Does
openness or outward-orientation affect economic growth? and, (4) How does human
capital enter the analysis?
Although a large proportion of the studies examined here, relate government,

openness and human capital to economic growth, rather than TFP, we use these studies
as a basis for exploring how these same factors may help provide a better understanding
of TFP.
This chapter is organised as follows. Section 2.2 begins by presenting the

different interpretations of TFP, then reviews several growth accounting studies that
examine the contribution of FA and TFP growth in economic growth, particularly of the
East Asian economies. It also illustrates the different approaches to the study of TFP
and its determinants, and explains why we opt to exclude human capital from the
underlying production function in calculating our measure of TFP for the succeeding
analysis. Section 2.3 analyses the links between government and economic growth,
12
while Section 2.4 analyses the links between openness and economic growth. Section
2.5 briefly discusses the main issues regarding the role of human capital in the HPAEs
growth experience, and Section 2.6 presents the conclusions of this chapter.
2.2
What is TFP?
The concept of TFP and the idea that anything else not explained by traditional
inputs of production, i.e., land, labour, natural resources and physical capital, falls under
TFP, has been discussed in the literature on economic growth from the early 1930s
(Griliches, 1995), yet many analysts remain uncertain as to what it really is. Lipsey and
Carlaw (2001) summarise the different interpretations of TFP in the literature into the
following:
a)
TFP as a measure of all improvements in technology and increases in

efficiency over long periods of time. This interpretation is what they
refer to as the conventional view.
b)
TFP as a measure of externalities and other free gifts associated with

economic growth.
c)
TFP as a measure of our ignorance, of what we do not know.
Although Lipsey and Carlaw subscribe more closely to the second interpretation
of TFP in view of their study of long-run technological change, they, nonetheless,
acknowledge that, TFP clearly means different things to different informed observers
(p.3).
In this study, we do not attempt to come up with a new interpretation of TFP.

Instead, we attempt to better understand what TFP is by identifying factors that
influence it, particularly within the experience of the HPAEs. We do this by first
reviewing the literature on growth accounting, then examining different approaches to
the study of TFP and its determinants.
13
2.2.1
FA and TFP growth
Growth accounting provides a structured framework for assessing the

importance of the relative contributions of labour and capital inputs and TFP growth.3,4
Basically, it involves decomposing the growth rate of output into the contributions of
the growth of inputs (labour and capital) and the growth of TFP.
The growth
accounting equation can be derived from a typical constant-returns-to-scale CobbDouglas production function:5
Yt = At K t Lt
(1 )
(2.1)
where,
Yt = level of output at time t
At = an index for the level of TFP at time t
K t = level of physical capital inputs at time t
Lt = level of labour inputs at time t
and and (1-) are parameters that measure the respective elasticities of capital
and labour with respect to output.
From the preceding function the implied growth rate of output is:
Y& A&
K&
L&
= + + (1 )
Y A
K
L
(2.2)
While growth accounting is a useful tool in determining the relative contributions of FA and TFP
growth, it does not test for any statistical significance of the relationship between output growth and any
of the factor inputs, nor of TFP; neither does it provide any test for any causal relationship.
In discussing the growth accounting methodology, the measure of TFP essentially refers to
improvements in technology and efficiency, for simplicity of exposition.
5
In this example, technology is assumed to be Hicks-neutral (output augmenting) rather than Harrodneutral (labour augmenting) for simplicity. If the production function assumes technology to be Harrodneutral, such that Y = K(AL)1-, the corresponding share of the growth of TFP in the derived growth
accounting equation would be equal to (1-), and the rest of the equation would be similar.
14
where,
Y&
= growth rate of output
Y
A&
= TFP growth
A
K&
= contribution of the growth rate of capital stock
K
&
(1 ) L = contribution of the growth rate of the labour force
L
In most growth accounting exercises, is simply approximated by capitals
share in output or GDP, taken from national accounts statistics. A few studies (e.g. Kim
and Lau, 1994; Drysdale and Huang, 1997) derive from regressions.
It is important to note that in the above specification, technical progress is

assumed to be totally disembodied; that is, all technical progress is assumed to be net
of the effects of factor accumulation (Dowling and Summers, 1998). However, it is also
possible to adjust equation (2.1) in order to account for embodied technological
progress, which could be of either the labour-augmenting or capital-augmenting form.
In the growth accounting literature, adjustments to the basic growth accounting equation
are usually made by incorporating a labour quality variable (e.g. Collins and Bosworth,
1996) or by introducing a separate human capital variable (e.g. Kim and Lau, 1995).
Table 2.1 summarises the main results of a number of growth accounting studies
that have included the HPAEs in the sample. Kim and Lau (1994, 1995), Young (1995),
Collins and Bosworth (1996), and Dowling and Summers (1998), are among those who
view East Asian growth as largely due to massive accumulation of labour and capital
rather than growth in TFP. On the other hand, Kawai (1994), King and Levine (1994),
Sarel (1997), and Drysdale and Huang (1997) are among those who argue that, while
labour and capital accumulation is necessary to foster growth, the unusually high rate
15
Table 2.1
Author
Selected growth accounting studies on East Asia

Dataset
Main conclusionsa
Shares in output growth (%)b

capital
labour
TFP
King and
Levine
(1994)
104 non-oil
exporting
countries for
the period
1950-88
104 non-oil
capital accumulation
countries:
accounts for only a small
fraction of income growth 40
(p. 280)
Kawai
(1994)
7 HPAEs
(excluding
Japan), plus
China and
the
Philippines
(Phil.), 7
Latin
American
countries;
over the
period 197090
both capital accumulation

and TFP growth are
important in explaining
growth in developing
countries
4 NICs and 5
OECD
(France,
West
Germany,
Japan, UK,
US); over
different
periods
ranging
between
1948 and
1990
capital accumulation in the

NICs accounts for between
48 and 72% of the NICs
economic growth, and is
therefore the most
important source of growth
for the NICs
Kim and
Lau (1994)
n.a.
n.a.
Asia : 44
Latin
America:
61
Asia: 33
Latin
America:
59
Asia : 23
Latin
America:
-20
NICsc:
77
OECDc:
40
NICsc:
23
OECDc:
5
NICsc:
0
OECDc:
55
NICsd:
60
OECDd:
35
NICsd:
18
OECDd:
6
NICsd:
22
OECDd:
59
NICse:
75
OECDe:
60
NICse:
16
OECDe:
6
NICse:
9
OECDe:
33
differences in TFP growth

are the major element
explaining the contrast in
the growth experience of
Latin American and Asian
countries
for the OECD countries,

technical progress is the
most important source of
growth, accounting for
between 46 and 71% of
their growth
results are generally the
same regardless of whether
factor shares used were
estimated from a metaproduction function or
obtained from a
conventional TFP
formula
continued over page
16
Table 2.1 (continued)
Author
Kim and
Lau (1995)
Young
(1995)
Collins and
Bosworth
(1996)
Dataset
4 NICs and
5 OECD
(France,
West
Germany,
Japan, UK,
US); over
the period
1970-74,
1989-91
4 NICs;
1960s to
1990s
88 countries
including 6
HPAEs
(without
Japan and
Hong
Kong), plus
the Phil.),
for the
period 196094
Main conclusionsa

capital
labour
TFP
without human capital:

contribution of physical
capital to overall growth of
the NICs is most important
and contribution of TFP
growth nil; for OECD
countries, TFP growth is
most important
Without
H:
NICs:
77
OECD:
40
NICs:
23
OECD:
5
NICs:
0
OECD:
55
with human capital:

contribution of physical
capital to overall growth in
NICs is still most important
and contribution of TFP
growth is still nil; for
OECD countries,
contribution of TFP growth
is still most important
with H:
capital:
NICs:
68
OECD:
35
NICs:
20
OECD:
6
NICs:
0
OECD:
54
average 3
NICs (w/o
Sing.)
37
average 3
NICs (w/o
Sing.)
25
Human:
NICs:
12
OECD:
4
average 3
productivity in the four
NICs, except Singapore, is NICs (w/o
Sing.)
not particularly low but
compared to productivity in 38
industrial countries, is not
Singapore:
unusually high either;
65
factor accumulation on the
other hand, has been
extraordinarily high from
the postwar period for all
four NICs
East Asian (6 HPAEs plus
the Phil.) growth is due
largely to factor
accumulation, not TFP
growth
East
Asia: 60
Singapore: Singapore
2
33
East
Asia: 14f
East
Asia: 26
continued over page
17
Author
Sarel (1997)
Dataset
5 ASEAN
countries
(Indonesia,
Malaysia,
Phil.,
Singapore,
Thailand)
and US for
the period
1978-96
Dowling and 6 HPAEs

(excluding
Summers
Japan and
(1998)
Hong
Kong), plus
China, India,
and the
Phil., for the
period 196195, and over
3 subperiods:
1962-75;
1976-85 and
1986-95
Main conclusionsa
careful calculation of factor

shares is imperative since
even small variations
would significantly affect
TFP calculations

capital
labour
TFP
ASEAN
(except
Phil.): 48
ASEAN
(except
Phil.): 14
ASEAN
(except
Phil.): 38
US: 44
US: 28
US: 27
n.a.
n.a.
TFP growth significantly

contributes to growth rate
of output per capita in
Indonesia, Malaysia,
Singapore and Thailand
during the period 1978-96.
Singapore appears to be
the best performer in the
region in terms of TFP (p.
23).
estimates of TFP are highly
sensitive to sample period,
size of capital share and
the stage of economic
growth and development of
a particular country
TFP is an important factor
in East Asian growth, but
during periods of high
growth, role of capital is
equally if not more
important (p.183) in
explaining sustained
growth in East Asia
1961-95
average
TFP
growth
estimate:g
1.6
( = 0.4)
1.9
( = 0.35)
2.25
( = 0.3)
disembodied technical
progress is more important
in explaining growth in the
early stages of
development
continued over page
of total factor productivity growth in the East Asian economies was, nonetheless, an
equally important contributing factor to the rapid growth of their respective GDPs per
capita.
18
Author
Drysdale
and Huang
(1997)
Dataset
Main conclusionsa
APEC
member
countriesh,
including
the HPAEs
over the
period 195090
productivity growth in the

East Asian economies, in
general, has been a
significant factor in their
rapid growth since the
1950s; contributing over a
third of per capita income
growth in Japan, Hong
Kong, and Taiwan; but the
contribution of productivity
growth to output growth in
Singapore (10%), Malaysia
(-8%) and the Philippines
(4%), was less significant
capital accumulation has
also been an equally
important factor

capital
labour
TFP
Japan: 54
Japan: 12
Japan: 34
China: 53
China: 40
China: 7
NICs: 37
NICs: 36
NICs: 27
Indonesia: Indonesia: Indonesia:

31
30
39
Malaysia:
60
Malaysia:
48
Malaysia:
Phil.: 49
Phil.: 47
Phil.: 4
Thailand:
29
Thailand:
42
Thailand:
29
-8
Notes:
n.a. indicates that the value is not available.
a
Results of the growth accounting studies do not appear to be sensitive to whether it is overall
growth (i.e. GDP growth) that is explained (e.g. Kawai, 1994; Kim and Lau, 1994, 1995;
Young, 1995; Dowling and Summers, 1998) or output per capita growth (e.g. King and Levine,
1994; Collins and Bosworth, 1996; Sarel, 1997; Drysdale and Huang, 1997).
b
The percentage contributions were calculated by averaging the respective shares for each
group of countries and over a common time period as specified.
c
The shares were computed based on parameter estimates from a meta-production function
(an approach used to estimate the relationship between aggregate output and inputs including
technical progress, from inter-country data), when restrictions on the rates of capital
augmentation are imposed, or parameter estimates obtained from a conventional TFP
formula.
d
The shares were computed based on parameter estimates when restrictions on the rates of
capital augmentation are imposed.
e
The shares were computed based on parameter estimates from a conventional TFP formula.
f
Labour contribution figures were adjusted for labour quality in terms of education per worker.
g
These figures are estimated growth rates of TFP calculated using King and Levines (1994)
capital stock data and Summers-Heston (1991) data, and not percentage contributions to
growth.
h
APEC member countries: Australia, New Zealand, Canada, US, Japan, China, Hong Kong,
South Korea, Taiwan, Indonesia, Malaysia, Philippines, Singapore, Thailand, Chile, Mexico,
Papua New Guinea.
19
The conflicting results, i.e. in terms of whether FA or TFP growth, is more
important, evident from the summary table, point to three important issues to consider in
the study of the determinants of economic growth. First, growth accounting appears to
be very useful in determining the contributions of FA and TFP growth in overall
economic growth, but the results are quite sensitive to the estimate of used, and the
level of development of the economy.
Secondly, intuitively, we would expect that the final measures of FA or TFP

growth would be sensitive to the way in which human capital is accounted for. That is,
if human capital is identified as a factor of production, either as a separate factor or in
terms of its effect on labour and/or physical capital, then its effect on the growth rate of
output is counted as factor accumulation. If, on the other hand, human capital is not
included as a factor of production, then any increase in human capital would form part
of the residual (along with any other omitted factor), and be interpreted as TFP growth.
It is important to note that among the studies cited in the table, only Kim and Lau (1995)
and Collins and Bosworth (1996) have made provisions for examining the contribution
of human capital to economic growth in a growth accounting framework.6
Gemmell (1996) shows four different ways of including human capital in

economic growth models (p.10):
(1)
in an extended Cobb-Douglas production function, such that growth in

output is a function, inter alia, of the rate of growth of human capital;
(2)
in Mankiw et al.s (1992) augmented Solow model that includes

accumulation of both physical and human capital;
(3)
in Romer (1990), where the stock and growth of human capital are
hypothesised to affect growth indirectly via its impact on physical capital
investment
The contribution of human capital to economic growth, has nonetheless featured significantly in many
other empirical studies of endogenous growth, e.g., Romer (1986, 1990), Becker et al. (1990), Tallman
and Wang (1994), Ahn and Hemmings (2000) and Temple (2001) among many others.
20
(4)
as a factor that affects growth by facilitating improvements in TFP via

imported technology
Although there are different ways of incorporating human capital in economic

growth models, there is no definite consensus regarding which of these is the most
appropriate. As Islam (1995) writes, Empirical work has so far clearly established that
human capital plays a very important role in the growth process. However, the question
that remains still unresolved is, In What Exact Way? (p.1154).
Mankiw, et al. (1992) show that production functions with human capital as an
additional factor of production fit better than those without human capital based on their
cross-section regressions. However, Islam (1995) finds otherwise, i.e., human capital
does not directly influence output growth, rather human capital influences growth
through its effect on TFP.7 Knowles (1994) suggests that intuitively, human capital,
measured in terms of both education and health status should be treated as labouraugmenting rather than a separate factor of production. Knowles and Owen (1997)
however, investigate whether incorporating human capital as labour augmenting or as a
separate factor of production makes any difference empirically. They find that it does
not.
In this study, we opt to exclude human capital from the underlying production
function in deriving TFP levels, following the findings of Benhabib and Spiegel (1994)
that cast doubt on the traditional role in the development process given to human capital
as another input of production.
Their findings, which are robust to a number of
alternative specifications and data sources, suggest that human capital plays a role in
determining productivity, rather than entering on its own as a factor of production.
The difference in the results obtained by Mankiw et al. (1992) and Islam (1995) may not be totally due
to the inclusion or exclusion of a human capital variable. It is also possible that their results are different
because the former use cross-sectional data, while the latter uses panel data. The inclusion of the time
dimension of the human capital variable in Islams panel data appears to outweigh its cross-sectional
characteristics.
21
Nelson and Phelps (1966) also argue that investments in human capital,
measured in terms of education, promote growth not only or mainly by enhancing the
capacity of labour, but rather by accelerating the rate at which individuals adopt and
adapt to new technologies. This results in the speeding of the process of technological
diffusion throughout the economy. They propose two specific models of the process of
technological diffusion and the role of education, which show that returns to education
are greater, the more technologically advanced the economy is. This suggests that
technological progress has implications for the optimal capital structure in the economy.
Another important point from their models is that if innovations (by more educated
workers) produce positive externalities, then education also yields externalities.
Because of these inter-connections between education and growth, Nelson and Phelps
suggest that the usual straightforward way of including human capital (using some index
of educational attainment) as an additional factor of production would represent a gross
misspecification of the productive process, specifically of the relationship between
education and the dynamics of production.
Benavot (1989) also argues that the main contribution of human capital, which is
generally and often crudely measured in terms of school enrolment ratios, to the growth
process, is to increase the level of cognitive skills possessed by the labour force, and as a
consequence, to improve productivity in general (p.15); and therefore should be
attributed to TFP. Hence for the purposes of this study, we likewise propose that, any
effect of human capital on overall economic growth is actually transmitted via its effects
on TFP.
Lastly, after determining the relative importance of TFP in the economic growth
of the HPAEs, we then should analyse the factors that could drive TFP growth. In this
regard, the next sections examine three basic approaches to studying the determinants of
TFP. We then review the role of governments and openness, which have featured in
many of the empirical studies on East Asian growth, as among the significant factors
that affect the HPAEs economic growth. The role of human capital in the economic
22
growth of the HPAEs is also discussed briefly before the concluding section of this
chapter.
2.2.2
TFP and its determinants within a single growth equation
One way of estimating TFP is to estimate a typical production function with the
usual factor inputs, labour and capital, but with additional factors, e.g., openness and
government policy indicators, that are assumed to affect growth via TFP. Knight et al.
(1993) analyse growth in this manner.
They estimate a Cobb-Douglas production
function similar to Mankiw, et al. (1992):
Yt = K t H t ( At Lt )
where the variables and parameters have the usual meanings.
(2.3)
In this particular
specification however, A, the labour-augmenting factor which reflects the level of

technology and efficiency (TFP) in the economy, is assumed to be dependent on the
degree of openness of the economy (F) (with elasticity f), and the level of government
fixed investment (P) (with elasticity p), and grows according to the following function:
At = A0 e gt F f P
(2.4)
where, g is the exogenous rate of technological progress.
Knight et al. (1993) believe that technological improvements are often absorbed
into the domestic economy via imported capital goods, and that efficiency of the
productive sector is influenced largely by the level of government investment in the
economy.
For these reasons, they consider inclusion of openness and government
variables crucial in the empirical analysis of economic growth.
23
Skipping the mechanics behind the derivation of the equations for the behaviour
of the other variables around the steady-state, Knight et al. (1993) estimate an economic
growth equation (for a panel data set) of the general form:8
ln y i ,t ln y i ,t 1 = 1 ln( ni ,t + g + ) + 2 ln s ki ,t + 3 ln s hi + 4 ln Fi
+ 5 ln Pi + ln y i ,t 1 + t + i + i ,t
(2.5)
where the variables have the usual interpretations (lower case letters represent growth
rates), t and i are time-specific and country-specific effects, and 1, , 5 and are
parameters to be estimated. Results of their estimation show, among other things, that
overall economic efficiency (TFP), is significantly and positively influenced by the
degree of openness to foreign trade and by the level of government investment in the
economy. Note however that within this particular framework, it is difficult to actually
separate the direct effects of F and P on overall growth from the indirect effects via
TFP.
2.2.3
TFP from growth accounting
Another approach to studying the determinants of TFP, is the one taken by

Collins and Bosworth (1996), who derive a measure of TFP from a growth accounting
framework that includes a human capital index, H.9 This human capital index measures
the benefits or returns to schooling, and is assumed to be embodied in the workers. The
underlying production equation takes the form,
Q = AK ( HL )
(1 )
(2.6)
In their derivation however, the growth rates of F and P are assumed to be zero for every time period, in
such a way that these growth rates drop out of the implied equations for estimation (see Knowles and
Owen, 1997 for an analysis of Knight et al.s derivations).
Hall and Jones (1999) approach the study of TFP in a similar manner, but opt to derive TFP from a
levels accounting exercise rather than through growth accounting. This will be discussed further in the
later sections of this review.
24
The contribution of capital is assumed to be uniform at 0.35 for their entire
sample of 88 countries. Growth of output per worker (q/l) is then decomposed into the
contributions of the growth of the physical capital per worker (k/l), growth of education
per worker (h), and the growth of total factor productivity (a):
q l = (k l ) + (1 )h + a
(2.7)
Assuming that education is the most important factor that affects worker
characteristics, Collins and Bosworth (1996) construct a labour quality index that
measures the quality of the workforce based on a 7% return to each additional year of
schooling. Educational attainment data are obtained from Barro and Lee (1994),
supplemented by data from Nehru et al. (1995), while the estimated returns to schooling
are based on data from Psacharopoulos (1994). A relatively low estimate of returns to
schooling (i.e., 7% against 12%) is used in calculating the labour quality index to
counter the effects resulting from omitted variable biases which Collins and Bosworth
observe tend to lead to overestimation of the returns to schooling, especially in the
developing countries. Also, to avoid ex-ante conclusions regarding low productivity
growth in East Asian countries, a lower rate of return is preferred, because a higher rate
of return to schooling results in a smaller residual (or TFP estimate) in the East Asian
countries (Collins and Bosworth, 1996). Results of their growth accounting exercise
show that TFP growth in all developing countries, including East Asia, has been
modest. Moreover, if the assumed return to schooling is increased from 7 to 12% (a
fundamental component of the human capital index, H), the contribution of TFP is
reduced further by an average of 0.4 of a percentage point.
In order to further analyse the ways in which East Asian countries differ from the
rest of the world, the resulting TFP growth rate is then regressed on some basic
measures of initial conditions (i.e., initial level of income, life expectancy and years of
schooling, as indicators of health and education); the external environment (i.e., the
mean and standard deviation of the annual change in each countrys terms of trade), and
25
some regional dummy variables.10 Results of the TFP regression indicate that initial
conditions significantly affect variations in TFP growth, and that regional effects only
marginally explain variations in TFP growth.
Since government intervention has often been cited as a major influence in the
success of the East Asian economies, Collins and Bosworth (1996), extend their analysis
of the determinants of TFP by regressing the changes in TFP (obtained from their initial
growth accounting exercise), on various macroeconomic and trade policy indicators
(i.e., budget balance, change in terms of trade, standard deviation of terms of trade,
standard deviation of real exchange rate, and a Sachs and Warner (1995) openness
indicator).11 Of the macroeconomic and trade policy indicators, only real exchange rate
stability appears to be consistently significantly associated with TFP growth. Collins
and Bosworth (1996) conclude therefore that these policies affect growth through
increasing capital accumulation, rather than through TFP growth.
Robertson (1999) analyses the relationship between the results of growth

accounting and the predictions of the neo-classical growth model. He finds that growth
accounting results are misleading because growth accounting implicitly treats capital
accumulation as an exogenous variable, which means that the dependence of capital
accumulation on productivity growth, a prediction of the neo-classical growth model, is
ignored. Hence, labour and (total factor) productivity contributions to growth will
always be underestimated within the growth accounting framework.12
10
Collins and Bosworth (1996) also use the contribution from capital accumulation as a dependent
variable in a similar regression.
11
Again, Collins and Bosworth (1996) also run a similar regression with the contribution from capital
accumulation as dependent variable.
12
Barro (1999) provides a general critique of the growth accounting framework, and shows how dual
approaches to growth accounting (i.e., where the residual is computed from changes in factor prices rather
than factor quantities), and allowing for spill-over effects and increasing returns, taxes and different types
26
2.2.4
TFP as residuals from a production function
In other empirical studies, such as Thomas and Wang (1993), Coe et al. (1997),
and Miller and Upadhyay (2000), TFP estimates are derived from typical production
functions, either with or without a human capital variable; and then regressed on other
variables, e.g. trade/openness and government policy indicators that are perceived to
affect TFP growth.
Thomas and Wang (1993) study the relationship between government policies
and productivity growth across 68 developing countries from the 1960s to 1980s, with
particular emphasis on East Asian countries, according to two alternative models, one of
which is based on the neo-classical Solow-Swan model of growth, while the other is
based on endogenous growth theory. Only their model based on the neo-classical
approach is discussed here because of its particular emphasis on TFP. The residual
from the estimated Cobb-Douglas production function is considered to be TFP, and TFP
is assumed to be a function of government policies, a regional dummy variable for East
Asia, and interaction terms between East Asia and government policies. From this,
Thomas and Wang propose two ways to estimate the model. The first method involves
estimating the production function with all the factors of production and policy variables
(similar to the approach of Knight et al., 1993):
ln Yit = a i 0 + j Pijt + j Pijt D + D +

j
ln X it + it
(2.8)
X = k ,l ,h ,e
where,
Yit
is the output of country i in year t;
Xit
are production inputs: physical capital (k), labour (l), land (h), and
educational capital (e), all of which are in log differences;
of factor inputs in the growth accounting equation, can affect the interpretation of the Solow residual as a
measure of technological progress.
27
Pijt
is policy indicator j for country i in year t;
is a regional dummy variable for East Asia;
PijtD
are interaction terms between East Asia and government policies

and distortions; and,
it
is a white noise error term.
The second method involves estimating the production function first. Then the
residual, i.e., TFP growth (TFPG), is regressed on government policy variables, an East
Asian dummy variable, and some interaction terms between East Asia and policy
variables:
TFPG it = 0 + j Pijt + j Pijt D + D + it

j
(2.9)
Results of their estimations generally indicate that government policies,

measured in terms of two indices that measure macroeconomic stability (Index 1) and
government expenditures (Index 2), significantly affect productivity growth. Index 1 is
positively and significantly associated with productivity growth, while Index 2, up to a
certain level, is positively and significantly associated with productivity growth, but
after a threshold level, becomes negatively associated with growth.13
Coe et al. (1997), following recent theories of economic growth and international
trade, derive empirical equations that relate a countrys TFP to foreign research and
development (R&D), capital stock, imports of machinery and equipment relative to
GDP (which measures the degree of openness to trade) and the secondary school
13
Further details on Thomas and Wang (1993) are provided in Section 2.3 Government and Economic
Growth.
28
enrolment ratio.14 TFP is first calculated as the log of output (GDP) minus the weighted
average of the logs of capital and labour inputs. The weights used are the factor shares.
Capitals share is set to 0.4. Two linear equations that link TFP to the foreign R&D
capital stock, the degree of openness and educational attainment are then specified and
estimated.
The equations differ only in the inclusion in the second equation of
interaction terms between foreign R&D capital stocks and import shares, and foreign
R&D capital stocks and secondary school enrolment rates:
log TFPit = i0 + iS log S it + iM M it + iE E it + itT Tt + it
log TFPit = i0 + iS log S it + iM M it + iE E it

+ iSM M it log S it + iSE E it log S it + itT Tt + it
(2.10)
(2.11)
where,
TFPit is total factor productivity in country i and time t;
is
are country-specific parameters;
is the foreign R&D capital stock;
is the share of machinery and equipment imports from industrial

countries in the GDP of each developing country;
is the secondary school enrolment ratio;
is a time trend; and
is a white noise error term.
The total factor productivity estimation results generally indicate that TFP is
positively and significantly related to the foreign R&D capital stock, and that a more
14
In an earlier paper, Coe and Helpman (1995), also derive similar empirical equations that look at the
relationship between TFP and R&D capital stocks and openness to trade. The main differences in the
later version however are: the inclusion of a proxy for human capital; the disregard of domestic R&D
capital stocks (since these were found to be in general, negligible in developing countries); and the focus
on imports of machinery and equipment to measure the degree of openness to trade (since imports of other
goods and services have been determined not to directly affect productivity). Chen, et al. (1997) examine
the econometric foundations of Coe and Helpman (1995) using panel data, and basically concur with the
latters results.
29
open and more educated economy also leads to higher productivity. Results of the
empirical estimation of the TFP equation with the interaction terms also show that
foreign R&D capital stock only affects developing country productivity through its
interaction with machinery and equipment imports (Coe et al., 1997, p. 135). That is, a
country with higher imports of machinery and equipment derives more productivity
gains from foreign R&D, and vice versa. However, there is insufficient evidence to
conclude that a more educated work force affects marginal productivity benefits from
foreign R&D.15
Miller and Upadhyay (2000) estimate two measures of TFP, one from a
production function with a human capital variable, and the other from a production
function without a human capital variable. Expressing the production functions in terms
of output per worker and taking natural logarithms, the basic estimated equation,
adjusted in consideration of the time-specific and country-specific characteristics of the
panel data set used, is:
ln y = ln A + ln k + ln h + ( + + 1) ln L + i timei + t
(2.12)
i =1
where, y is real GDP per worker, L is the number of workers, k is physical capital stock
per worker, h is human capital measured as the average years of schooling per adult, A is
an index of total factor productivity, , and are parameters to be estimated, and,
timei (i = 1, , 6) are time dummy variables ( for the six time periods each consisting of
five-year observations over 1960-1989).
Equation (2.12) is estimated with = 0
imposed, and with 0, and results show that inclusion of a human capital variable
only slightly improves the fit of the regression equations.
15
Engelbrecht (1998) criticises the results of Coe et al. and re-evaluates their earlier findings.
Engelbrecht, using more appropriate measures of human capital, and a model which simultaneously
incorporates both embodied and disembodied research and development (R&D) spillovers, concludes that
human capital is positively associated with other knowledge spillovers that impact on TFP growth.
30
Miller and Upadhyay (2000) then examine the effects on growth when
interactions between human capital and physical capital, and human capital and the
labour force are allowed for. Their results show that the coefficient of the interaction
term for human capital and physical capital is positive and significant, indicating that
human capital does affect the output elasticity of physical capital. On the other hand,
the coefficient on the interaction term for human capital and the labour force is
insignificant.
These results suggest a positive link between both types of capital
(physical and human), but no link between human capital and the labour force.
After estimating TFP from the preceding equations (equation (2.12) with = 0
imposed, and with 0) Miller and Upadhyay (2000) test for significant determinants
of TFP, by estimating the following equation with trade and openness variables, which
are assumed to directly affect TFP:
ln tfp = a1 + a 2 ln H + a 3 ln x + a 4 ln tot + a 5 ln pd
+ a 6 ln (1 + ) + a 7 ln x + a 8 ln tot + a 9 ln pd
(2.13)
+ a10 ln + a10+ i timei +

i =1
where,
TFP
is total factor productivity derived from the estimated production

functions;
is the human capital stock;
is the ratio of exports to GDP;
tot
is the terms of trade;
pd
is the deviation of the local price from purchasing power parity;
is the inflation rate; and,
is the standard deviation of i (= x, tot, pd and ) over the five-year subperiods.
31
Time dummy variables are again included in the estimating equation in order to
capture fixed effects over time. In general, results of Miller and Upadhyays estimation
show that TFP growth is positively and significantly related to the trade/openness
variables regardless of whether TFP was derived from a production function with or
without a human capital variable. This suggests that a more open economy is associated
with higher TFP. When Miller and Upadhyay re-estimate their equation to include an
interaction term between human capital and openness (in terms of the ratio of exports to
GDP), they find the coefficient of the interaction term positive and significant, implying
that more human capital contributes to a larger effect of openness on TFP. However,
further analysis shows that the effect of human capital on TFP is initially negative, and
becomes positive only after a certain level of openness.
The preceding sections describe different approaches to studying TFP and its
determinants. Although each of the different approaches has its merits, the study of
TFP remains relatively ad hoc and, as Prescott (1998) concludes, a sound theory of TFP
is still needed. This, however, should not prevent us from persisting in our search for
the factors that drive TFP, and specific to the objectives of this study, help us explain
the economic success of the HPAEs.
The next three sections examine the role of
government, openness and human capital in the HPAEs growth experience.
2.3
Government and Economic Growth
Analysis of the role of governments in the East Asian miracle is presented in

three parts. First, we discuss the areas of debate regarding the role of governments in
the East Asian economies. Next we briefly look into the different ways in which
governments have actually intervened in the HPAEs; and finally, we analyse some of the
main empirical works in this area.
32
2.3.1
To intervene or not to intervene
The specific role that governments played in the economic success of the HPAEs
has been the subject of much debate. At one extreme of the controversy lies the freemarket theory, according to which, the efficiency of the price mechanism in the HPAEs
was responsible for rapid economic growth. The governments role in the East Asian
economies was limited primarily to the provision of a stable macroeconomic
environment that allowed markets to function efficiently (e.g. Chen, 1979; Friedman
and Friedman, 1980). Chen further asserts that large capital investments, which were
necessary for the rapid growth of the East Asian economies (particularly, Japan, Hong
Kong, Singapore,16 South Korea and Taiwan), were a result of high rates of domestic
savings, which were achieved without direct government intervention (p.184).
At the other end of the debate, lies the theory of governed-markets, that supports
the idea that the economic success of the HPAEs is largely attributable to active
government intervention in the economy (e.g. Hasan, 1976 and Amsden, 1989 for South
Korea; Wade, 1990 for Taiwan; Johnson, 1982 for Japan; Jomo and Gomez, 1997 and
Snodgrass, 1998 for Malaysia). According to Wade (1990, p. 27):
The corporatist and authoritarian political arrangements of East Asia

have provided the basis for market guidance ... effected by augmenting
the supply of investible resources, spreading or socializing the risks
attached to long-term investment, and steering the allocation of
investment by methods which combine government and
entrepreneurial preferences.
At the centre of the two extreme theories, is the market-enhancing view, which
suggests that markets and government are not mutually exclusive substitutes (Aoki et
al., 1997).
According to this view, the governments of East Asia, recognising that
markets fail, did not hesitate to step into the economic scene, but only insofar as
interventions remained market-friendly.
16
The World Development Report 1991
Note that foreign investment was also an important factor in Singapores economic growth. Domestic
savings was, therefore, not the only source of investment funds.
33
(World Bank, 1991), defines a market friendly government as one which encourages
investment in human capital, promotes fair competition in the private sector, supports
integration with the global economy, and maintains a stable macroeconomic
environment.
While, this view recognises certain advantages of government
intervention, the World Bank (1993a) nonetheless points out that, beyond these marketfriendly functions, governments are likely to do more harm than good (p. 84).
2.3.2
Government intervention in the HPAEs
In the face of the arguments put forward above, it is quite difficult to determine
which theory best explains the role of governments in East Asian development. This is
because the policy mix, strategy and degree of government intervention have been
different across the East Asian economies, and over time. For instance, Hong Kong and
Singapore are more free-market oriented than the other HPAEs. Government policy in
Hong Kong has always been seen as the closest approximation to laissez faire in the
world economy. Hong Kongs policy of positive non-interventionism means that the
government does not attempt to plan the allocation of resources available to the private
sector or to frustrate the operation of market forces, but it still assumes full
responsibility for providing the private sector with an economic environment and the
infrastructure conducive to industrial development (Root, 1996).
Singapore has been described as the most striking example of an economy which
relies largely on the market, particularly world market forces, yet, the general orientation
of the economy remains significantly influenced by well-defined government policies.
With limited physical resources, Singapore was forced to depend heavily on foreign
direct investment (FDI) and multinational corporations (MNCs) to industrialise, and to
attract foreign investors, the government intervened substantially in the labour market,
particularly in establishing wage-setting policies (Leipziger and Thomas, 1993).
The Thai governments economic policies have also emphasised the role of
market forces and minimum government intervention (intervention was basically limited
34
to the provision of infrastructure) to spur economic growth (Warr and Nidhiprabha,
1996).
By contrast, Taiwans and South Koreas governments are considered to be the
most interventionist - picking winners and deliberately interfering in market
operations, distorting relative prices - in order to stimulate economic activity (Amsden,
1989). In South Korea, following the example of Japan, the government encouraged
private enterprises to invest in prioritised industries (labour-intensive export industries
in the 1960s and heavy and chemical industries (HCIs) in the 1970s) by providing
favourable credit terms, special tax concessions and infrastructural support. The entire
banking sector was also nationalised, so that, effectively, all foreign loans were obtained
by the government, and it was the government which decided to whom to extend the
credit, based on export performance. By the early 1980s however, South Korea suffered
from considerable economic difficulties resulting primarily from the unprofitability of
the HCIs, so the government decided to reconsider its policies. The HCI drive was
abandoned, trade was liberalised, restrictions in the financial sector were reduced
(commercial banks were de-nationalised), and the system of providing cheap credit to
officially endorsed projects was also abolished (EIU, 1988).
The Taiwanese government also identified strategic industries but opted to
intervene through direct investment and establishment of a network of state-owned
enterprises (SOEs) which have been considered as instruments for the big push in
Taiwans growth (Wade, 1990; Macintyre, 1994). To illustrate, the percentage share of
Taiwans SOE output in GDP was 13.5% over the period 1978-1980, compared to the
Asian average of only 8.0% (Wade, 1990).
The Malaysian and Indonesian governments also intervened substantially in their

respective economies. Under Malaysias New Economic Policy (NEP), ethnic Malays
were supported through subsidies, employment quotas, scholarships and investment
licensing to enable them to penetrate the industrial sector (Balassa, 1991).
The
government also set up SOEs to provide employment for the Malays and extended
preferential credit to Malay businesses. In Indonesia, domestic financing of the budget
35
through debt or money creation was banned from 1966 (Bhattacharya and Pangestu,
1997), and high earnings from the export of petroleum were used instead to finance
public investments under an ISI (import-substituting industrialisation) development
strategy.
However, Leipziger and Thomas (1993) observed that government
interventions in Malaysia and Indonesia were not quite successful prior to the 1980s.
But when Malaysia suffered from recession in the mid-1980s, brought about by
inefficiencies in the implementation of the NEP, and Indonesias export earnings from
oil started to decline in 1982, resulting in severe balance of payments problems in
Indonesia,17 both governments re-directed their policies towards less intervention, and
since then a marked improvement in the economic performance of these two countries
has been evident (Islam and Chowdhury, 1997).
2.3.3
Analysing the empirical evidence
The ways in which governments in East Asia have intervened show that
governments can try to affect the rate of growth of their respective economies through
various policies. These policies may be directed at improving resource allocation and
overall factor productivity (e.g. through preferential tax treatments for firms in the more
efficient sectors of the economy), encouraging the flow of investment (domestic and
foreign) into the country (e.g. through careful management of interest rates and foreign
exchange rates), promoting technical progress through public and private investments in
research and development (e.g. by guaranteeing patents, royalties, copyrights, etc.), and
promoting the development of human capital (e.g. through educational subsidies, setting
of educational standards to ensure quality of education provided in state and private
educational institutions and better access to these institutions). However, there is still
no consensus in the results of empirical work that analyses the relationship between
government interventions and economic growth (see Table 2.2).
17
Earnings from oil exports make up more than 70% of Indonesias central government revenues (Mishra,
1995). Indonesias oil industry is practically dominated by Pertamina, a state-owned oil and natural gas
mining company, which has production-sharing arrangements with foreign oil companies.
36
Table 2.2
Selected empirical work on government and economic growth
Author
Dataset
Methodology
Main Results
Kormendi
and Meguire
(1985)
47
developed
and
developing
countries;
over the
period
1950-77
OLS
while government
spending appears to be
positively correlated with
economic growth, the
relationship is not
statistically significant
115
developed
and
developing
countries;
over the
period
1960-80
OLS based on a two-sector

production function framework
(government and non-government
sector) in which output in each sector
depends on labour and capital inputs
and that output from the government
sector has an externality effect on the
output of the other sector
Ram (1986)
dependent variable: mean annual rate

of growth of aggregate real gross
domestic or national product
explanatory variables: initial per
capita income; mean population
growth rate; standard deviations of
real output growth and money supply
shocks; mean of growth of money
supply, government spending to
output ratio, export to output ratio,
rate of inflation; and mean investment
to income ratio
2 basic OLS regressions: crosssection regressions over different

periods (1960-80; 1960-70; and,
1970-80), to bring out possible
structural variations over time; and
OLS on time series data for the 115
countries
in almost all countries,

overall impact of
government size on
growth is positive
output in the government
sector has positive
externality effects on the
rest of the economy:
effects are stronger in the
1970s than in the 1960s;
also stronger among the
LDCs
results are similar across
countries and across time
dependent variable: annual GDP

growth rate
explanatory variables: growth rates of
investment to output ratio and
government to output ratio;
population growth rate
continued over page
37
Author
Dataset
Methodology
Main Results
Barro
(1991)
98
developed
and
developing
countries;
over the
period
1960-85
OLS
growth rate of real GDP

per capita is negatively
related to the ratio of
government consumption
to real GDP
2 data sets:
OLS
crosssection data
set of 100
countries
over 197088
dependent variable: growth rate of

GDP per capita
Easterly and
Rebelo
(1993)
annual
time-series
data for 28
countries
from 18701988
dependent variable: annual average

growth rate of real GDP per capita
explanatory variables: initial values
of real GDP per capita; investment
(total, private domestic and
government domestic) to output
ratios, ratio of public domestic
investment to total domestic
investment; ratio of real government
consumption (exclusive of education
and defense) to real GDP; fertility
rate; population growth rate; various
school enrollment rates; adult literacy
rates; measures for political
instability; dummy variables for type
of economic system; measures of
market distortions; and regional
dummies.
explanatory (fiscal) variables:

marginal tax rates; government
revenues; government expenditures
(with and without defense and
education); transfers; social security
contributions; investment on
transportation and communication
public investments in
transportation and
communication and
governments budget
surplus are consistently
correlated with growth
ratio of government
revenue to GDP
increases with per capita
income
link between other fiscal
variables and growth is
statistically fragile,
depending heavily on the
choice of fiscal variables
included in the particular
regression equation
(especially in the crosssection regressions)
continued over page
38
Author
Dataset
Methodology
Main Results
Barro and
Lee (1994)
2 data sets:
SUR (seemingly unrelated

regression) estimation to allow for
country-specific random effects that
are correlated over time, and
assuming errors for different
countries (either in same time period
or across the two time periods) are
uncorrelated
size of government;
government-induced
market distortions and
political instability are
negatively correlated
with growth
85
countries
over the
period
1965-75
95
countries
over the
period
1975-85
Kocherlakota and Yi
(1996)
annual
time-series
data from
1917-1988
for the US
dependent variable: real GDP per

capita growth rate
explanatory variables: initial level of
state variables (initial per capita
GDP; initial stock of educational
human capital; initial stock of human
capital); control or environmental
variables (investment to GDP ratio;
government expenditure to GDP
ratio; black market premium; measure
of political instability (no. of
revolutions per year))
OLS
dependent variable: GNP growth rate
explanatory variables: lags of GNP
growth rates, marginal tax rates and 7
policy variables (2 tax measures, 4
public physical capital investments,
growth rate of M2)
only investments in nonmilitary equipment

capital and non-military
structural capital
positively and
significantly affect longrun growth
continued over page

Based on the standard neoclassical growth framework, Barro (1991) and Barro
and Lee (1994), in separate studies of a cross section of countries, including the HPAEs,
over the period 1960 to 1985, show that growth is negatively related to three
government-related variables: government expenditures (less expenditures on defence
and education); political instability, measured in terms of the number of coups,
revolutions and political assassinations; and government-induced price distortions,
39
Author
Dataset
Methodology
Main Results
Thomas and
Wang
(1996)
pooled
crosscountry and
time-series
data set
consisting
of 68
countries
(10 East
Asian
countries
and 58 low
and middle
income
countries)
over the
period
1960-90
2 growth models are analysed: the

first is based on a neoclassical
Cobb-Douglas production function,
where the residual is considered to
be the growth rate of TFP which is
assumed to be a function of
government policies, a regional
dummy for East Asia, and
interaction terms between East Asia
and the government policies; while
the second follows an endogenous
growth model, wherein economic
growth is a function of government
policies through their influence on
the rate of return to capital and the
rate of factor accumulation
trade openness and

macroeconomic stability
are positively and
significantly correlated
with economic and
productivity growth
92 countries
over the
period
1960-89
both exogenous and endogenous

growth is modelled under a single
stochastic growth model, that shows
growth rate of per capita output to
be mean stationary if growth is
exogenous, and difference stationary
if growth is endogenous
Evans
(1997)
assuming that different types of

government policies affect growth
differently, two indices which
distinguish between distortions and
interventions are constructed: Index
1 which proxies for distortions,
refers to trade and price
irregularities and macroeconomic
instability; while Index 2 which
proxies for interventions, encompass
various types of government
expenditures
government expenditures
positively affect growth
up to a particular level,
but thereafter, affect
growth negatively
overall trade, price and
macroeconomic policies
in the East Asian
economies are associated
with better and stronger
economic growth,
compared with countries
in other regions
economic growth rate is

not significantly related
to the share of
government consumption
in GDP
economic growth rate is
also found to be mean
stationary, which is
evidence for the
exogeneity of growth
continued over page
40
Author
Dataset
Methodology
Main Results
de la Fuente
(1997)
19 OECD
countries
over the
period 1970
-1995 for the
panel
estimations;
over the
period 1960 1995 for the
review of the
main features
of the data
estimates a series of regressions of

the growth rate of income per capita
and the private investment rate on
different combinations of
government revenue and
expenditure indicators and a set of
non-fiscal control variables
government expenditures
in general, have a
negative impact on
income levels and growth
rates, both directly and
through their effect on
private investment
also, an extension of MRWs (1992)

model is estimated, modified to
allow for direct estimates of the
coefficients of a reduced-form
aggregate production function
but, one component of

government expenditure,
public investment,
appears to have a positive
impact on productivity
growth, but is subject to
sharply diminishing
returns
85 countries
over the
period
1975-1990
cross-country and panel data

analyses are carried out
economic freedom affects

growth through an effect
on TFP and investment
Dawson
(1998)
data are generally averages over the

period for the cross-country
analysis; for the panel analysis, data
are divided into three five-year subperiods (1976-1980, 1981-1985 and
1986-1990)
an extension of MRWs (1992)
endogenous growth model is
estimated, including three freedom
indices: economic, political and
civil
political and civil

freedoms are not
statistically significant
influences on growth, but
some evidence that these
freedoms are statistically
significant influences on
investment
continued over page

measured in terms of the black-market premium on foreign exchange.18 In both studies,
government expenditures are believed to decrease the level of savings, and
consequently, through the distortionary effects of high taxes and government
18
Barro and Lee (1994) choose to use the black market exchange premium to proxy for price distortions
because it is objectively measurable and available for a wide range of countries.
41
Author
Dataset
Methodology
Main Results
Kneller et
al. (1999)
panel of 22
OECD
countries,
data
aggregated
into 5-year
averages
over the
period
1970-95
endogenous growth model, which

includes different elements of
government
distortionary taxation
reduces growth, while
non-distortionary taxation
does not
5 different forms of panel data

estimator are considered: pooled
OLS, one-way (country dummies)
fixed (by OLS) and random (by
GLS) and two-way (country and
time effects) fixed and random
effects models
productive government
expenditure enhances
growth, while nonproductive expenditure
does not
government variables are classified

into four main types: distortionary
taxation (taxes on income and profit,
social security contributions, taxes
on payroll and manpower, property
taxes), non distortionary taxation
(taxes on domestic goods and
services), productive expenditures
(general public service expenditure,
defence, education, health, housing
and transport and communication
expenditures), and unproductive
expenditures, social security and
welfare expenditure, expenditures
on recreation and economic
services), plus other revenues and
other expenditures
Bleaney et
al. (2001)
panel of 22
OECD
countries,
two types of
data: annual
data and
periodaveraged
data over
the period
1970-95
as above, but estimations include

lagged values of the fiscal variables
to allow for a long-run effect of
fiscal policy on growth
main results as above, but

also finds that period
averaging and static panel
methods do not fully
capture the long-run fiscal
effects on growth
42
programmes that discourage private-sector activities, decrease the level of output per
worker. Price distortions and political instability influence growth rates through their
(negative) effects on the level of investment.
Ram (1986) however, concludes otherwise in his study of 115 countries

(including Hong Kong, Indonesia, South Korea and Singapore) over the period 1960 to
1980. His analysis is based on a two-sector production function, in which output in
each of the two sectors of the economy: government and non-government, is assumed to
depend on labour and capital inputs; and that output in the government sector has spillover or externality effects on the output of the other (non-government) sector. His main
result indicates that the over-all impact of government size (again, measured in terms of
government expenditures) on economic growth is positive, and argues that a larger
government can be a powerful engine of growth because of governments capacity to
reduce conflict between private and public interests, prevent foreign exploitation,
generate increases in productive investment, and provide a socially optimal direction
for growth and development (Ram, 1986 p. 191).
He further concludes that
government size also exhibits positive externality effects on the rest of the economy and
that the effects are particularly strong during the 1970s; while factor productivity in the
government sector is higher in the 1960s.19
These results were consistent across
countries and across time, but the positive effects of government appeared to be stronger
in the poorer countries, including South Korea and Singapore (which were considered
poor countries in the early 1960s20).
On the other hand, Kormendi and Meguire (1985) using a simple production
function framework, also find a positive relationship between government expenditures
and growth in their study of a cross-section of 47 countries, but find the effect to be
statistically insignificant, leading them to conclude that there is no significant
19
Alexander et al. (1996) however, note that Rams results should be treated with extreme caution
because of some problems associated with Rams assumption that the government sectors marginal factor
productivity and externality effects are likely to have the same (positive) sign.
20
In 1960, Korea had a GDP per capita equivalent to 898 (in 1985 international prices), which was
lower than GDP per capita in Bangladesh (939), Somalia (1100) and Sri Lanka (1253). Singapores GDP
per capita in 1960 was slightly higher at 1658 (Summers and Heston, 1991).
43
relationship between government and economic growth.
Evans (1997) concludes
likewise in his study of growth based on a stochastic growth model that analyses
exogenous and endogenous growth within a single framework. He tests the stationarity
of GDP growth rates using data from 92 countries from Summers and Heston (1991)
Penn-World Table (PWT) 5.6 and hypothesises that if growth rates are mean stationary,
then growth is exogenous, but if they are stationary in differences, then growth is
endogenous.
Moreover, if growth is found to be endogenous, then any variable
affecting growth should likewise be difference stationary.
Using the share of
government consumption in GDP as his explanatory variable, he suggests that growth

rates would be negatively related to government consumption if, and only if, growth is
endogenous (p. 209). However, the results of his study show that the growth rate is not
significantly related to the share of government consumption in total output, and is
exogenous.
One possible reason why the results of the aforementioned studies have been
conflicting is related to the way the government intervention variable is measured in
each study. Much of the empirical work on the relationship between government and
economic growth makes use of data on government expenditures, as a proportion of
either total output or private investment, as a measure of government size or degree of
government intervention in the country (e.g. Ram (1986), Kormendi and Meguire
(1985), Barro (1991), Barro and Lee (1994), Evans (1997)).
If one simply looks at aggregate government expenditures (measured in terms of

the average share of government expenditure in GDP) across different country
groupings over the period 1960-1990, it appears that HPAE governments intervened
much less21 (Figure 2.1), yet achieved higher growth rates, compared to the respective
averages in other country groupings. However, the relationship between government
and growth is not this simple.
21
Average percentage share of government expenditure in GDP over the period 1960-1990, is 12.49% in
the HPAEs, 13.74% in OECD countries, 15.67% in Latin America and 20.89% in the rest of the world
(calculated from Summers and Heston (1991) data).
44
gov't expenditure share in gdp (%)
Figure 2.1
Average share of government expenditure in GDP (%)

(in 1985 international prices)
24
22
20
18
country groupings
16
rest of the world
14
oecd
12
latin america
10
1960
1964
1968
1972
1976
1980
1984
1988
1962
1966
1970
1974
1978
1982
1986
hpaes
year
Source: Summers and Heston (1991) PWT 5.6
Government expenditures alone can be poor representations of actual effects of

government intervention on various sectors of the economy. As can be gleaned from the
discussions in earlier sections, government interventions take several forms (e.g. tax
concessions; setting of interest rates, foreign exchange rates, wage rates; subsidising
education; direct public infrastructural investments, etc.).
Another reason why government expenditure is a poor measure of government

intervention relates to the way government expenditures are aggregated. It is desirable
to disaggregate government expenditures by type of expenditure (e.g. defence/military
expenditures, education, social welfare, or public capital investments), because each
type of government expenditure can affect different sectors of the economy differently.
For instance, expenditures on education can lead to higher productivity growth, but
public capital investments could crowd out private consumption and investment.
Kocherlakota and Yi (1996), in a study of how public structural capital

investment affects the long-run economic growth of the United States, disaggregate US
public physical capital investments into four classes: non-military equipment capital,
non-military structural capital; military equipment capital; and military structural
capital. Using OLS, GNP is regressed on lagged values of GNP and seven policy
45
variables: marginal tax rates, tariff rates, the four classes of public physical capital
investments, and the growth rate of M2, which proxied for monetary policy. They find
that only non-military equipment capital and non-military structural capital significantly
positively affect the long-run economic growth of the United States.
Similarly, Kneller et al. (1999, pp. 173 and 177) classify government variables
into distortionary taxes (those that affect investment decisions with respect to physical
and/or
human
capital),
non-distortionary
taxes
(those
that
do
not
affect
saving/investment decisions), productive expenditure (those that have a direct effect on

growth), and non-productive expenditure (those that do not affect the steady-state of
growth). They find that, When financed by some combination of non-distortionary
taxation and non-productive expenditure, an increase in productive expenditures
significantly enhances growth, and an increase in distortionary taxation significantly
reduces growth (p. 188). Recently, Bleaney et al. (2001) using panels of annual and
period averaged data (five-year averages), show that averaging the data into five-year
periods (as commonly done in empirical studies) does not fully capture the long-run
responses, but estimating the endogenous growth model using either annual data or
period averages produces similar results, i.e., strong support for the fiscal effects on
growth. Another important contribution of Kneller et al. and Bleaney et al. is how they
consider the effects of both government spending and taxation on economic growth, and
not just either one in isolation.
Another study that has attempted to include more government variables (e.g.
disaggregated expenditures on education, defence, social services, etc., marginal tax
rates, tariff rates, growth rates of money supply, government revenues, inflation rates) to
capture the effects of various aspects of government policy, is that of Easterly and
Rebelo (1993).
However, despite including more government variables into the
regression equations, they have not produced very conclusive results either.22 They use
two data sets, one covering 100 countries over the period 1970 and 1988, and the other,
22
Kneller et al. (1999) note that such non-robust results may in part be due to lack of proper attention to
the linear restriction implied by the government budget constraint (p. 176).
46
28 countries over a longer historical period, 1870-1988, and find a strong association
between the level of development and different indicators of fiscal structure23 in both
data sets. Growth is found to be positively associated with public investments on
transportation and communication, but total public investment appears to have no
significant effect on growth. Effects of the different tax variables used are difficult to
establish, because the coefficients are often negative, yet statistically insignificant.
Also, estimation problems (e.g. strong correlations between the fiscal variables used)
ensue when more than one fiscal variable is included on the right-hand side of the
regression equation.
Dawson (1998) studies the empirical relationship between institutions,

investment and growth using indices on economic, civil and political freedom, taken
from Gwartney et al. (1996), and various issues of Gastils Freedom in the World.
Dawsons results indicate that economic freedom has a positive effect on growth, but
the effects are rather indirect, as institutions affect growth either through TFP or
investment. The relationship between growth and either political or civil freedom is not
statistically significant; however, there appears to be some mixed evidence that these
freedom indices affect investment. Dawsons study is another attempt at studying the
alternative channels through which government can affect growth. However, the use of
aggregate indices for economic, civil and political freedom, implicitly assumes that each
of the specific components of the indices affects growth and/or TFP similarly. For
instance, Gwartney et al.s economic freedom index is composed of 17 components that
include measures of money and inflation, government operations and regulations,
takings and discriminatory taxation and restrictions on international exchange.
By
aggregating these 17 measures into one summary economic freedom index, specific
effects of any of these measures on growth and/or TFP can no longer be identified.
The study by Thomas and Wang (1996) also attempts to isolate the specific
effects of government intervention on economic growth. In a study of 68 countries over
23
The fiscal variables used were: tax rates, government revenues and expenditures (with and without
defense and education), transfers, social security and public investments on transportation and
communication.
47
the period 1960 to 1990, Thomas and Wang investigate the association between
government intervention and distortions with TFP growth in East Asia, where
interventions have been common, and compare this to the situation in other developing
countries. Interventions in this study are defined as government actions measured in
terms of public-sector investment and government spending as shares of GDP (e.g.,
fixed capital formation, expenditures on health and education, etc.), while distortions
refer to trade and price irregularities and macroeconomic instability. Thomas and Wang
hypothesise that distortions (measured in terms of Index 124) negatively affect growth
performance, but the effects of government intervention (measured in terms of Index
225) are likely to be more ambiguous, since this would depend not only on the extent of
intervention made, but also on the nature and quality of intervention. Again, indices are
used, and we have similar reservations regarding the use of indices as in Dawson
(1998). However, a particular feature of this study is the distinction made between
intervention and distortion, which is made on the premise that, while government
interventions and distortions are interrelated, they do not coincide.
Government
interventions do not always lead to distortions, and not all distortions are a result of
government intervention.
Thomas
and
Wangs
results
show
positive
relationship
between
macroeconomic stability (Index 1), and economic growth, but they find no significant
linear association between intervention (Index 2), particularly as measured by
government expenditures, and growth. Nonetheless, a non-linear relationship between
expenditures and growth is evident, with government expenditures positively affecting
GDP and productivity growth up to a point, but then starting to negatively affect GDP
and productivity growth at higher levels.
The point Thomas and Wang raise in relation to the nature and quality of
intervention merits further consideration. To illustrate, the path of Latin American and
24
Index 1 which measures macroeconomic stability (and openness) was based on data on trade openness,
real interest rates, foreign exchange premium in the parallel market, and inflation rate.
25
Index 2 was based on average data for public sector investment, government fixed capital formation,
subsidies, productive expenditures and total government expenditures.
48
East Asian (except in the case of Hong Kong) industrialisation has often been regarded
as state-led since the 1950s, but the objectives, instruments and consequences of
government intervention in these countries, have been varied. A number of studies (e.g.
Hartlyn and Morley, 1986; Isuani, 1994; Maddison, 1995) suggest that the policy mix
adopted by the Latin American countries (e.g. import-substitution industrialisation (ISI),
excessive foreign borrowing, inattention to social services such as education, etc.) has
been misguided,26 and only led to slower economic growth, compared with the growth
in East Asian economies, whose governments adopted a policy mix more conducive
to economic growth (e.g. outward or export-oriented industrialisation (EOI), promotion
of human capital development, prudent public sector spending, etc.) (World Bank,
1993b).
Looking at the different strategies adopted by the HPAE governments and the
consequent effects on the economic growth of their respective countries, it appears
extremely important, in any attempt to study the relationship between government and
economic growth, to use better measures for the role of government than simply using
aggregate government expenditures. For instance, government expenditure should be
disaggregated to distinguish between positive intervention and distortion.
2.4
Openness and Economic Growth
The first part of this section examines the different avenues by which openness is
perceived to affect economic growth, while the second part analyses the empirical
evidence (or lack of it) in support of this.
26
Another stream of studies, e.g. Wiarda (1987) and Singh (1993), suggest that while the Latin American
governments made mistakes in their policy choices, there were other compelling reasons why their
economies did not fare as well as the economies in East Asia, e.g. the debt crisis, and the existence of
stronger populist movements in Latin America that undermined political stability in the region.
49
2.4.1
Establishing the link between openness and economic growth
The interest in the study of the link between openness and economic growth
stems from the documented inefficiencies of the inward-looking strategies popular in the
1950s and 1960s (Cline, 1987; Edwards, 1993). However, it was not until the late
1980s that trade liberalisation became a popular option for achieving economic growth.
Most likely, it was only after the world observed documented evidence of the slow
growth in the Latin American economies, which followed more protectionist and
restrictive trade policies, compared to the rapid growth of the more open East Asian
economies, that policy makers and economic advisors begin to take serious interest in
the effects of a freer trade policy on economic growth. Figure 2.2 compares the level of
openness, defined by Summers and Heston (1991) as exports plus imports as a share of
GDP, of selected East Asian and Latin American economies.27
Figure 2.2 shows that the East Asian nations had, on average, more open
economies compared to the Latin American nations during the period 1965-1990. Over
this same period, the average annual rate of growth of real GDP per capita of the East
Asian economies was about 6.7%, a rate much higher than the Latin American
economies average rate of per capita income growth of only 0.8% (Villegas, 1998).
Based on these figures, it appears that more open economies do tend to grow faster.
However, the implied positive correlation from these figures does not take into
consideration other variables that could affect either openness or economic growth;
hence the relationship between the two variables could be more complicated than the
simple correlations imply. Furthermore, as will be discussed later, the use of export and
import shares relative to GDP is not a very adequate measure of the actual level of
openness of a country.
27
East Asian economies: Hong Kong, Japan, South Korea, Taiwan and Thailand; Latin American
economies: Mexico, Argentina, Bolivia, Chile, Colombia and Venezuela. Openness figures are
unweighted regional averages.
50
Figure 2.2
Level of Openness: East Asian vs. Latin American economies

120
openness
100
80
60
region
40
east asia
20
1965
latin america
1969
1967
1973
1971
1977
1975
1981
1979
1985
1983
1989
1987
year
Source: Summers and Heston (1991) PWT 5.6
The most common argument in support of the hypothesis that openness affects
economic growth rests on the notion that openness/trade leads to improvements in total
factor productivity, as a result of either static or dynamic gains from trade.28 Static
gains from trade arise as a result of reallocating resources to the production of goods for
which the economy has a comparative advantage. By specialising in the production of
goods that the economy can produce at lower opportunity cost than other countries, the
economys consumption possibilities expand. That is, the country can consume outside
its production possibilities frontier (PPF). On the other hand, dynamic gains occur when
trade improves the production possibilities of an economy, i.e. an outward shift of the
PPF.
As Pack and Page (1994) point out, economies which are producing below the
international best-practice production function29 must invest in hardware and
software to move towards it. Foreign exchange generated by exports is therefore very
valuable to enable the country to access new and better technologies and move towards
best practice. A more open economy can also gain from economies of scale since the
28
29
See for example, Seguino (2000) and the references cited therein.
This is basically the production capacity at which further increases in output at given levels of inputs
cannot be achieved without the introduction of new techniques (Pack and Page, 1994, p. 201).
51
international market is much larger than the domestic economy (Emery, 1967; Balassa,
1978, 1988).
In addition to the static and dynamic gains, a country can also benefit from
competitive gains from trade. In order for an economy to maintain its position in the
international market, its export industries are pressured to keep production costs low,
operate in a more efficient manner, and maintain high quality export products, and/or
continuously improve products (Emery, 1967). This competitive pressure from the
international market also stimulates the flow of market innovations, managerial skills,
etc. to the exporting country, which are likely to induce total factor productivity growth.
These competitive pressures force the economy to operate on the PPF rather than inside
it.
Nonetheless, while there may be strong theoretical grounds supporting the

hypothesis that openness positively affects economic growth, primarily through TFP
growth, empirical research in this area has so far yielded conflicting results. Table 2.3
summarises a number of empirical studies on the relationship between openness and
economic growth.
2.4.2
Analysing the empirical evidence
As Table 2.3 indicates, most empirical results regarding the relationship between
openness and economic growth show that economic growth is positively
correlated with openness. However, a careful inspection of the studies in terms of the
variables considered and methodologies employed in each study brings out two main
areas of weakness in the general literature on openness and economic growth. The first
area relates to the seemingly arbitrary use of different proxies for openness or trade
orientation, while the second relates to the inability to deal convincingly with causality
issues, especially since most of the empirical studies have relied heavily on crosssectional data (Edwards, 1993).
52
As can be observed from Table 2.3, different authors used different measures to
proxy for openness. Emery (1967), Balassa (1978), Jung and Marshall (1987), and
Serletis (1992), use export growth rates as an indicator of a countrys level of openness
or trade orientation. Feder (1982) and Chow (1987) also use export growth rates but,
specifically, growth rates of manufactured exports. Ahmad et al. (1997) use the level of
exports per capita. Pack and Page (1994) and Young (1994) use the average share of
manufactured exports in total exports, the share of manufactured exports in GDP, and
Dollars (1990) index of openness.30 Further still, Edwards (1992), Taylor (1995),
Harrison (1996), and Gundlach (1997)use various measures of price distortions and
trade policy indicators (e.g. tariff averages, quotas, black market exchange rates, trade
reforms, disprotection of agriculture, etc.) in their analyses.
The common practice of using trade volumes or trade shares and the rate of
growth of exports as a proxy for openness is limited because exports are not necessarily
related to trade policy. A countrys trade regime can be heavily distorted, but still have
high export growth rates (Edwards, 1997). For instance, based on Dollars distortion
index, South Koreas economy is more heavily distorted than the economies of Brazil,
Mexico, the Philippines, India, Thailand and Turkey (Fishlow et al., 1994), but all have
lower export growth rates than South Korea.
However, even Dollars distortion index cannot escape criticism. Rodriguez and
Rodrik (1999) check the robustness of Dollars index and find that this index is
theoretically appropriate as a measure of trade restrictions only when three conditions
hold: (1) there are no export taxes or subsidies in use, (2) the Law of One Price holds,
and (3) there are no systematic differences in national price levels due to transport costs
and other geographic factors (Rodriguez and Rodrik, 1999, p. 12). Obviously these
three requirements do not always hold true in practice.
30
Dollars (1990) outward-orientation index is based on data on price levels of traded goods from
Summers and Heston (1988).
53
Table 2.3
Selected empirical work on openness and economic growth
Author
Dataset
Methodology
Results
Emery
(1967)
cross-country: 50
countries
multiple correlations and OLS

(ordinary least squares)
significant positive
correlation between export
growth rate and GNP per
capita growth rate
1 time period:
1953-63 average
dependent variable: GNP per

capita
explanatory variables:
exports, current account
earnings
Balassa
(1978)
Feder
(1983)
cross-country: 10
developing
countries
Spearman rank correlation

and OLS
2 time periods:
1960-66; 196673
domestic and foreign capital,
labour and export growth
(trade orientation is measured
in terms of export growth)
cross-country: 31
semiindustrialised
less developed
countries
OLS
1 time period:
1964-73 average
Jung and
Marshall
(1985)
cross-country: 37
developing
countries
time period:
annual from
1950-81
dependent variable : GNP
dependent variable: GDP

growth rate
explanatory variables: share
of investment in GDP, growth
of population (proxy for
labour growth), and export
growth multiplied by exports
share in GDP
Granger causality testsb and
F-tests
variables considered: annual
GNP growth rates and annual
export growth rate; lagged
values of GNP and export
growth rates
export growth has a greater

effect on the rate of economic
growth than do the
contributions of domestic and
foreign capital and labour
marginal factor productivities

in the export sector are higher
than in the non-export sector,
and shifting factors from a
low productivity sector (nonexports) to a high real
productivity sector (exports)
contributes significantly to
economic growtha
lack of support for hypothesis
that export growth causes
output growth
continued over page
54
Author
Dataset
Methodology
Results
Chow
(1987)
cross-country: 8
Asian and Latin
American NICs
Sims (1972) causality testc
strong causal relationship

between export growth and
industrial development; most
countries exhibit bidirectional causalities,
therefore export growth and
industrial development are
mutually beneficial and
reinforce each other
time period:
annual from
1960-80s
Serletis
(1992)
Edwards
(1992)
variables considered: output

of manufacturing industries,
exports of manufactured
goods
1 country:
Canadian
unit root and cointegration

tests, Granger causality test
time period:
annual from
1870-1985
variables considered: GNP,

exports and imports
cross-country:
30 developing
countries
OLS using basic data set and

alternative trade orientation
indicators and alternative data
sets to test for robustness of
results
1 time period:
1970-82 average
alternative data
sets
cross-country: 51
developed and
developing
countries
different time
periods: 1970-82
average; 1960-82
average
dependent variable: growth

rate of real GDP per capita
investment ratio, knowledge
gap, two intervention
indicatorsd, four openness
indicese; and 8 alternative
indicators of openness and
trade distortionsf
export growth has significant

positive effect on GNP
growth
there is a strong and robust

positive relationship between
trade orientation and
economic performance;
results also show evidence of
a catch-up effect, i.e.,
countries with a lower initial
income per capita will tend to
grow faster than other
countries; issues related to
causality are still open
continued over page
55
Author
Dataset
Methodology
Pack and
Page
(1994)
cross-country:
113 countries
cross-country regressions
1 time period:
1960-85 average
Results
exports, not openness, are

relevant for high rates of
dependent variable: - average
productivity change; countries
rate of growth of GDP per
that enter more intensively
capita
into international trade are
also able to converge more
average share of investment
rapidly towards higher
in GDP, a measure of
income status - move towards
educational attainment, rate of international best practice
growth of economically active
population, relative gap
between GDP per capita in
1960 and US per capita
income in 1960
other variables - 2 measures
of export performance: ave.
share of manufactured exports
in total exports and share of
manufactured exports in
GDP; and Dollars (1990)
index of openness
Young
(1994)
cross-country:
113 countries
(Pack and Page,
1994 dataset)
1 time period:
1970-85 average
cross-country regressions;
review and critique of Pack
and Page (1994)
dependent variable: growth of
GDP per capita
explanatory variables: initial
income (1960) primary
enrollment rate in 1960,
average share of investment
in GDP over the period 196085, average share of
manufactured exports
evidence in support of the

argument that manufacturing
exports positively affect
growth is sensitive to the
nature of manufacturing
exports to GDP data and the
sensitivity of the regressions
to the inclusion of the NICs
and the separation of initial
and final shares; rapid growth
of the NICs should not be
viewed as evidence of
potential dynamic gains from
outward-oriented policies
continued over page
56
Author
Dataset
Methodology
Results
Taylor
(1995)
pooled crosscountry dataset

as averages for
4 successive 5year periods:
207 countries, 4
time periods:
1970-74;197579;198084;1985-89
2SLS (2 stage least squares)

growth accounting regression
high investment rates were

the key to high growth rates
in the Asia-Pacific region,
which in turn are caused by
low distortions, among
others; much of the
investment in the developing
economies is in the form of
imported capital goods, so
high investment environment
might be more aptly termed
import-oriented growth
several data sets

(panel and
cross-country)
depending on
the availability
of openness
measures for
each country
rank correlations across

different openness measures;
OLS regressions based on pure
cross-country data and then on
panel data; Granger causality
tests between openness and
growth
found positive relationships

among seven openness
measures, but correlations
using panel data are more
statistically significant than
correlations using pure crosssectional data
dependent variable: GDP

growth rate
only one of seven openness

variables positively affects
growth using cross sectional
data; while six of seven
measures affect growth using
panel data
Harrison
(1996)
cross-country:
from 17 to 51
countries
time period:
annual from
1960-87 or
1978-88
dependent variable: growth

rate of GDP per capita
explanatory variables: initial
per capita GDP, initial years of
primary and secondary
schooling per person in the
population, 4 endogenous
control variables related to
factor accumulation, 10
environmental variables that
measure natural resource
endowments, demographic
structure, government policies,
price distortions and openness,
and financial intermediation
and monetary instability
explanatory variables: capital

stock, years of primary and
secondary education,
population, labour force,
arable land and technological
change; openness measures trade reform (1960-84), trade
reform (1979-88), black
market premium, Dollars
price distortion index, trade
shares, dis-protection of
agriculture
causality between openness

and growth was found to run
in both directions
continued over page
57
Author
Dataset
Methodology
Results
Gundlach
(1997)
cross-country:
13 open and
9 closed
developing
countries
cross-country regression
based on Mankiw, Romer and
Weil (1992), augmented by a
level dummy for openness
open developing countries

converge at a much higher
rate to their steady state than
closed developing countries
1 time period:
1980-92
average
Ahmad,
Harnhirun
and Yang
(1997)
cross-country: 5
ASEAN
countries
(Indonesia,
Malaysia,
Philippines,
Singapore,
Thailand)
time period:
annual from
1950-81
dependent variable: real GDP

per worker
investment in human and
physical capital, level dummy
for openness
unit root and cointegration
tests; error correction
modelling and final
prediction error (FPE)
method in selecting
appropriate lag length
Granger causality testsa
variable considered: annual
GDP per capita and level of
exports per capita
exports do not Granger cause

economic growth in any of
the countries
for Indonesia and Singapore,
economic growth causes
export growth
no evidence of causality in
either direction in the case of
Malaysia, the Philippines and
Thailand
Notes:
Alexander et al. (1996) question the reliability of Feders results due to his treatment of
inference regarding marginal factor productivity differentials.
b
Following Granger (1969), a variable x is said to Granger cause another variable y if current
y can be predicted better by using past values of x than by not doing so, using all other past
information from a given information set which includes x and y.
c
Sims (1972) concept of causality refers to the statistical analysis of past (lag) and future
(lead) relationships between economic time series, say, between y and x.
d
2 intervention indices, one using unscaled (homoskedastic), and the other, scaled
(heteroskedastic) with residuals, models to estimate trade flows.
e
4 openness indices, two overall indices derived from using scaled and unscaled trade models,
and the other two indices are for the manufacturing sector, derived from scaled and unscaled
trade models.
f
Alternative indicators of openness and trade distortions: average black market premium;
coefficient of variation of average black market premium; index of relative price distortions;
average import tariffs; average non-tariff barriers coverage; World Development Reports index
of trade distortions; index of effective rates of protection; and, World Bank index on outward
orientation.
a
58
Rodriguez and Rodrik provide a comprehensive assessment of the different
measures of openness used in the empirical literature on openness and growth. They
find that in many cases, the indicators used by researchers are problematic as measures
of openness, or are often highly correlated with other sources of poor economic
performance.
For instance, the Sachs-Warner (1995) openness indicator is a zero-one dummy,

which takes the value 0 if the economy was closed based on any of the following
conditions being present during 1970 1989: (1) average tariff rates are higher than
40%, (2) non-tariff barriers cover at least 40% of trade, (3) average black market
premium exceeds 20% during the 1970s or 1980s, (4) economic system is socialist, and
(5) state has a monopoly over major exports. However, this binary classification of
countries into either open or closed puts on equal footing countries that may actually
have different degrees of openness. Rodriguez and Rodrik (1999) also find that the
strength of the Sachs-Warner index derives mainly from two of the five conditions
identified earlier: the black market premium and the state monopoly of exports. These
two variables however, may proxy for a wide range of policies, which may not be
exclusively associated with trade policy. The existence of a sizeable black market
premia over long periods of time (over 20% during the 1970s or 1980s based on the
Sachs-Warner condition) for example, reflects a wide range of policy failures, only a
portion of which may be related to trade policy failures. This also implies that the use of
black market premia on its own, i.e., not as a component of some general index, is a
rather poor proxy for openness.
Rodriguez and Rodrik (1999) remain sceptical about the use of various
indicators of trade policy.
They conclude that despite the proliferation of these
indicators, we are yet to find one that stands out as the best indicator of openness. We
are inclined to agree with this finding.
59
Pritchett (1996) examines cross-country data for six trade policy stance
measures31 that have often been used in empirical studies as proxies for trade policy
orientation. Pritchett hypothesises that if correlations among the different empirical
proxies for openness can be found, then, these different proxies are capturing some
significant, well-understood aspect of countries trade orientation. However, results
from his study indicate that the various trade policy indicators are nearly completely
unrelated in the cross-country data set (p.326). This suggests that some countryspecific, time-persistent aspect of trade policy is not adequately measured and this is
why different dimensions of trade policy may have different effects on growth. Harrison
(1996), also examines the rank correlation of seven different measures of openness32
using pure cross-sectional data as well as panel data. While the results of her study
show positive relationships among the different measures in both cross-sectional and
panel data, the rank correlations using the pure cross-sectional data are not statistically
significant. The statistical significance of results from panel data suggests that timeseries properties of the variables under study, which are lacking in pure cross-sectional
data, may have important implications for the resulting estimated relationships among
the variables.
In order to test whether the identified relationships hold regardless of the time
period of the cross-sectional data, Edwards (1992) performs growth regressions using
cross-sectional data sets averaged over different time periods. Results of his study show
that the positive relationship between openness and growth holds, regardless of the time
period considered.
Berthlemy, et al. (1997) study the effects of human capital on economic growth
using a panel data set of 83 countries over six 5-year periods, from 1960-65 to 1985-90.
31
The six trade policy stance measures examined were: (1) structure adjusted trade intensity (SATI),
which is a trade flow outcome measure of policy openness based on the residuals from a trade intensity
regression, (2) Leamers (1988) openness index, (3) average tariffs, (4) non-tariff barrier frequency, (5)
price distortion, and, (6) Leamers trade distortion indices.
32
The seven openness measures examined were: trade reform (1960-84), trade reform (1979-88), black
market premium, price distortion, trade shares, and disprotection of agriculture.
60
They assume that human capital development depends on the degree of openness of the
economy, and that the more open the economy, the higher the level of human capital
development and the higher the consequent economic growth of the country.
Berthlemy, et al. (1997) determine that the positive association between openness and
economic growth cannot be confirmed when tested on a panel data set. However,
Harrison (1996) estimates the impact of several measures of openness on economic
growth, using pure cross-sectional data on the one hand, and panel data on the other, and
finds that openness is associated with economic growth regardless of the type of data
used.
Only the strength of the positive association between the various openness
measures and economic growth differs, with the results obtained from panel estimation
being more statistically significant than the results obtained using cross-sectional data.
These results demonstrate that inclusion of the time-series properties of the data
into the analyses affects the conclusions reached in the studies. Furthermore, without
time-series data, the direction of causation between openness and economic growth is
difficult to establish,33 and this is another important issue to consider. This is because
even though a positive correlation between openness and growth can be demonstrated,
this will not determine the causal relationship between the two variables.
Empirical studies on openness and economic growth that deal with the issue of
causality suggest that causality can run in either direction. However, it is the premise
that real export growth causes real GNP growth that is tested more often. Although
rarely tested, it is also possible for output growth to cause export growth. As domestic
industries grow, production capacity expands and with a small domestic economy, the
foreign market serves as an outlet for products that cannot be absorbed by the domestic
economy; hence, the level of exports could grow.
Aside from the direction of causality, Jung and Marshall (1985) also point out
the possibility of negative correlations. If causation from economic growth to exports is
33
This essentially hinges on the way causality is defined, which will be discussed more extensively in
Chapter 4.
61
determined, it is possible that the sign of the effect is negative; e.g. real growth brought
about by an increase in domestic consumer demand for primarily exportable
commodities could lead to a reduction in the export levels. On the other hand, if
causality can be established from exports to economic growth, the overall effect could
also be negative. For instance, inward foreign direct investment (FDI) for exportoriented industries could increase export levels, but if the FDI is subject to various trade
distortions (e.g., tariffs, over-valued exchange rates), overall output of the domestic
economy could decline.34
Results of causality tests on exports and economic growth are mixed. Jung and
Marshall (1985) perform Granger causality tests for the sign and effect of exports on
economic growth and vice versa for 37 developing countries, and find little evidence
that exports promote economic growth. Only in one of the HPAEs, Indonesia, was this
evident. Their results also show stronger evidence of causation from output to exports,
but causation is negative, i.e. output growth reduces the level of exports. Ahmad, et al.
(1997) on a study of five ASEAN countries (Malaysia, Indonesia, the Philippines,
Singapore and Thailand), find no evidence of causality from exports to growth, but find
reverse causation (i.e. from growth to exports) in the case of Indonesia and Singapore.
On the other hand, Chow (1987) in his study of the four newly industrialised countries,
finds strong support for bi-directional causality (based on Sims (1972) causality test)
between manufactured exports and economic growth. Similarly, Harrison (1996) using
different measures of openness, finds that causality between openness and growth runs
in both directions (except when the black market premium is used, where causality is
found to run only from openness to growth).
This section has highlighted two main areas of weakness in the general literature
on openness and economic growth: (1) the difficulty of establishing what is the most
appropriate measure of openness or trade orientation to use in the study; and (2) how to
adequately capture and evaluate the time-specific effects of the relevant variables and
34
If however, FDI is mainly for domestic industries, then this could still lead to an increase in the over-all
level of domestic output.
62
determine the actual direction of causality between openness and economic growth.
This study attempts to address these weaknesses, first, by examining the effect of
different measures of openness on economic growth via TFP; and, second, by using
different types of data sets (pure cross-sectional, time-series and panel (pooled) data
sets), to test the causal and long-run relationships between openness and growth via
TFP.
2.5
Human capital and growth in the HPAEs
Any reference material on the HPAEs economic growth experience will not fail
to mention the HPAEs impressive human capital development, particularly in terms of
investments in education. Here are a few examples:
In nearly all the rapidly growing East Asian economies, the growth
and transformation of systems of education and training during the
past three decades has been dramatic. Today, the cognitive skill
levels of secondary school graduates in some East Asian economies
are comparable to, or higher than, those of graduates in high-income
countries.
World Bank (1993a) p. 43
All of the HPAEs have invested heavily in education and, unlike many
other developing countries, have concentrated on primary and
secondary schooling The benefits of focusing on primary and
secondary education are substantial. The higher the (sic) enrollment
rates in primary and secondary education, the higher the growth in a
countrys per capita gross domestic product.
Campos and Root (1996) p. 56
There is a well-entrenched view that the superior performance of the

economies of Pacific Asia relative to other economies has to do with
their human resources in the form of an abundant supply of skilled
workers Perceptions on the role of human resources in economic
development bear the imprint of human capital theory which posits a
direct link between a persons endowment of education and
productivity.
Islam and Chowdhury (1997) p. 89
63
It might seem obvious that investment in education was one of the
secrets of Asian growth Presumably low levels of education result
in low levels of productivity, perpetuating poverty and discouraging
foreign investors, and, therefore increasing educational levels would
have the opposite effect.
Tipton, (1998) p. 343
Claims like those quoted above, have led many people to accept as true that
human capital is indeed an essential ingredient of the HPAEs remarkable economic
performance. However, not all are convinced. Hughes (1995) notes for instance, that
while education in the Philippines in the 1950s was most advanced compared to that of
other East and Southeast Asian countries, and remains at a high level at present, the
Philippines has nonetheless, remained an economic laggard. Indonesia and Thailand
have lagged behind the Philippines in terms of human capital development35 yet; both
Indonesia and Thailand have been two of the fastest growing economies in Asia, casting
doubts on the role of human capital in economic growth. Even the results of empirical
studies in this area have been mixed. Behrman (1990) reviews a number of studies on
rates of return to human resource investments (both country studies and cross-country
studies), and concludes that:
There seems to be some evidence about the association between

human resource investments and development, though there are some
outliers for which such an association does not appear to hold well
Behrman (1990), p.89
Behrman also advocates the need for more systematic research into this area,
with careful consideration of causality issues and exploration of possible interactions
between human resource investments and other factors.
As discussed in Section 2.2, there are several ways of incorporating human

capital in economic growth analyses. For our purposes, we hypothesise that human
capital affects growth essentially through TFP, and examine the hypothesised
35
For example, the Philippines had practically achieved universal primary education by the 1960s, while
Indonesia and Thailand did not achieve this until the late 1970s; also, the secondary school participation
rate in the Philippines was 68% in 1986, compared with only 41% and 29% for Indonesia and Thailand,
respectively (Pernia, 1993).
64
relationships (cointegrating or long-run relationships, causal relationships, etc.) not only
between TFP and human capital, but also among human capital and other factors
(government and openness), and also consider the possible effects of the interaction
between human capital and openness on TFP.
2.6
Conclusions
This chapter has examined three of the most controversial issues regarding the
economic growth experience of the HPAEs: (1) How important is TFP in helping
explain the economic growth of the HPAEs and how do we account for the
contributions of human capital in growth? (2) What role do governments play in
economic growth? and, (3) Does openness affect economic growth? A definitive
resolution to the above areas of concern is yet to be reached. While human capital has
often been considered an important ingredient in the economic growth experience of the
HPAEs, we find that empirical evidence has not been entirely convincing, and more
research into this area is warranted.
In subsequent chapters of this study, we attempt to address the weaknesses of

previous empirical studies on the long-run economic growth experience of the HPAEs
by focusing on TFP and its determinants.
We use different types of data sets in our
analyses in order to account for cross-country variations and inter-temporal dynamics,

which are potentially crucial in comparing the growth experience of the HPAEs and
other countries across time. Although we lack a sound TFP theory within which to base
our analyses, we, nonetheless, offer a structured empirical framework within which the
hypothesised determinants of TFP (e.g. government, openness and human capital) can
be analysed, using appropriate estimation techniques, complemented with diagnostics to
test the robustness and reliability of the results.
65
~ CHAPTER THREE ~
Explaining Differences in Total Factor Productivity Across Countries
3.1
Introduction
In the previous chapter, we reviewed how economic growth is essentially the

result of factor accumulation (FA) and total factor productivity (TFP) growth.
However, growth through FA, is not sustainable in the long-run. Hence, in order to
obtain a better understanding of economic growth in the long-run, we focus instead on
TFP, which has been argued to be the crucial determinant of long-run economic growth,
and attempt to identify the factors that affect it. In this chapter, we hypothesise that
government intervention, the degree of openness of an economy and human capital are
the key factors that affect TFP, and that TFP differs across countries because of
differences in these factors.
As mentioned in the introductory chapter, this thesis also aims to provide

information on the robustness of the effects of a selected set of variables on TFP in a
manner consistent with current empirical practice. We therefore conduct our analyses
using different types of data: cross-sectional, time-series, and pooled cross-sectional and
time-series data, beginning in this chapter with our cross-sectional analysis. Crosssectional analyses have become an important starting point for discussions about
economic growth (see for example Romer, 1986 and Lucas, 1988). Within the context
of TFP studies, one aim of running cross-country regressions is to use the variation in
TFP across different units (i.e., countries) to understand its determinants.
Our analysis in this chapter is directed at examining the levels of TFP instead of
TFP growth rates because levels capture differences in the long-run relationships among
the variables of interest (i.e., TFP and its determinants).
Once the variables are
differenced, we lose valuable long-run information about the between-country level

relationships (Hendry, 1995).
As Baltagi and Griffin (1983) also show, between-
country variation can more accurately reflect long-run responses and, by contrast, timeseries variation presents only short-run responses. Furthermore, recent developments in
66
the study of the role of technological change in economic growth (e.g. Temple, 1999a)
suggest that countries eventually grow at a common, steady rate in the long-run because
of technology transfers. Based on this, it is more appropriate to study levels rather than
growth rates.
We acknowledge that one of the primary objections to cross-country studies

relates to the assumption that different countries (regardless of how very different they
are, e.g. Sierra Leone and Singapore) are drawn from a common surface, i.e., the
problem of parameter heterogeneity. A very simple test to determine how closely the
countries fall on a common surface is by looking at the regression R2 (Temple, 1999a).
typically, an of over 0.20 in cross-sectional regressions is indicative of some
explanatory powers, despite heterogeneity problems, among others (Green, 2000). We
also check for outliers and influential observations in our cross-section regressions to
help render our findings more robust.
Our analysis of TFP differs from previous empirical work in several ways.
First, in calculating an alternative measure of TFP levels, we consider how the
conversion of data in international prices may affect the TFP levels we calculate, as
well as the results of TFP regressions.
We also exclude human capital from the
production function used to calculate the TFP levels.
Second, we use different
measures of government intervention to examine the role of governments in

encouraging (or deterring) TFP growth in the economy. This is an improvement over
much of the previous empirical work that focused on government expenditures alone as
a proxy for intervention (see Chapter 2 Section 2.3). Third, diagnostic tests, which
appear to be lacking in much of the previous empirical work, are performed to examine
the reliability and robustness of our results.
The concept of levels accounting, as set up by Hall and Jones (1999) is

introduced in Section 3.2, and from this, we derive an alternative measure of TFP
levels. Section 3.3 examines correlations among the different measures of government,
openness and TFP. Section 3.4 examines whether or not our hypothesised determinants
are indeed significant influences on TFP, and Section 3.5 presents robustness and
67
diagnostic tests to verify our results. Section 3.6 concludes and provides a summary of
our findings.
3.2
Estimating TFP levels
In this section we introduce Hall and Jones levels accounting methodology, and
then discuss how we derive our alternative TFP measure.
3.2.1
Hall and Jones (1999) TFP levels
The approach we use in this study to determine the specific factors that affect
TFP, and why TFP levels vary across countries, is based on Hall and Jones (1999)
(hereafter HJ) levels accounting study of productivity across a cross-section of 127
countries. HJs levels accounting equation is derived from a typical Cobb-Douglas
constant-returns-to-scale production function, with labour augmented by a human
capital variable, such that:
Yi = K i ( Ai H i )
(3.1)
where the variables and parameters have the usual meanings, except H, which is human
capital-augmented labour, rather than just human capital.
HJ obtain their output per worker data from the Penn World Table 5.6 of
Summers and Heston (1991). They also construct physical capital stock data using a
perpetual inventory method using investment data also from the Penn World Table 5.6.
Using Mincer-based estimates of human capital, HJ express human capitalaugmented labour as:36
36
Mincer (1974) regresses the log of workers wages on the workers years of schooling and experience,
and finds that an additional year of schooling increases a workers productivity proportionally by the
derivative '(E), which is the return to schooling, estimated from the regression.
68
H i = e ( Ei ) Li
(3.2)
where Ei reflects years of schooling and (E) is a measure of the efficiency of a unit of
labour with respect to E years of schooling.
Re-writing equation (3.1) in terms of output per worker, y Y/L, and with
information on and , HJs levels accounting equation is expressed as:
K
y i = i
Yi
(1 )
hi Ai
(3.3)
where, h H/L, is human capital per worker, and A is the measure of total factor
productivity.
The preceding equation shows how the level of output per worker in each
country can be explained in terms of the contributions of the capital-output ratio
(following Mankiw et al., 1992 and Klenow and Rodriguez-Clare, 1997), human capital
per worker and TFP.
HJ assume = 1/3, which has often been used in growth accounting studies (e.g.
Maddison, 1987; Collins and Bosworth, 1996; Dowling and Summers, 1998). For the
returns to schooling, HJ use Psacharopoulos (1994) survey of returns-to-schooling
estimates for a wide range of countries, and assume = 13.4% for the first four years of
education, 10.1% for the next four years, and 6.8% beyond the eighth year of education.
Average educational attainment, measured in 1985 for the population aged 25 and over,
is taken from Barro and Lee (1993). HJ calculate A for a cross-section of 127 countries
for the year 1988.37
37
See http://www.stanford.edu/ chadj/HallJones400.asc for their complete dataset.
69
3.2.2
An alternative measure of TFP levels
Hall and Jones (1999) establish a very practical method of deriving TFP levels.
However, for our purposes, we opt to derive a new set of TFP levels following Hall and
Jones (1999) levels accounting methodology (we also assume = 1/3), but excluding
human capital from the underlying production function equation, i.e.,
Yi = K i ( Ai Li )
(3.4)
where the variables and parameters have the usual meanings. Re-writing equation (3.4)
in terms of output per worker, y Y/L, and again, following MRW (1992) in
decomposing y in terms of the capital-output ratio, the new levels accounting equation is
expressed as:
(1 )
K
yi = i
Yi
Ai
(3.5)
As in HJ, the measure of TFP is A in equation (3.5).
As discussed in Chapter 2 (Section 2.2.1 pp. 15-17), the inclusion of a human

capital variable in the underlying production function affects the final estimates of TFP.
Briefly, we note that including human capital as a factor of production would understate
the contribution of TFP in overall output growth. We also propose that any effect of
human capital on overall output growth is actually transmitted via its effect on TFP, and
we opt to keep it from the underlying production function in deriving TFP levels. In
constructing our alternative measure of TFP levels, we make use of data from HJ, but
exclude human capital in our calculations, as specified in equation (3.5).
We note that the estimates of TFP from this levels accounting exercise will be
sensitive to the assumed and also to the functional form of the levels accounting
equation. However, an assumed equal to 1/3 is consistent with previous growth
70
accounting studies (as mentioned earlier),38 and as Hall and Jones (1999) show, results
based on the Cobb-Douglas formulation and alternative results based on production
functions that are not restricted to Cobb-Douglas, are very similar.
3.3
Correlation analyses
Our analysis of the determinants of TFP begins by examining the correlations

between TFP and the different measures of government, openness and human capital, as
well as correlations across the different government and openness measures.
As
mentioned in the introduction to this chapter, we hypothesise that TFP differs across
countries due to differences in government intervention, the degree of openness of an
economy and human capital.
The motivations for the correlation analyses are: first, to determine whether the
different government and openness measures are associated with TFP and/or growth,
and, second, to determine whether the different government and openness measures are
correlated with each other and, if so, how highly correlated they are. If the correlations
among the different measures are not very high, then including them in the estimated
TFP equation may provide more information regarding the relationships being studied.
Note however that the correlation analyses are merely exploratory in nature and do not
control for the effects of other potentially relevant variables.
3.3.1
Data
Two measures of TFP are used in the preliminary correlation analysis. The first
is our alternative measure of TFP levels for 1988 (i.e., without human capital in the
production function), that we derive using data from Hall and Jones (1999).
The
second TFP measure is from Islam (1995), which is the residual from estimating a
typical Cobb-Douglas production function (without a human capital variable) within a
panel data framework. The correlations between various government and openness
38
In Chapter 4, we examine more closely the different TFP estimates obtained from assuming different
levels of for the HPAEs.
71
measures and measures of output levels and growth (e.g. real GDP, real GDP per
worker, growth rates of real GDP and real GDP per worker) are also examined. The
different measures of governments involvement in the economy and indicators for the
level of openness/trade are presented in Table 3.1.
Note that the Barro and Sala-i-Martin (1995) data are essentially from Barro and
Lee (1994), but using data from Summers and Hestons (1991) Penn World Table
(PWT) version 5.5. Barro (1991) uses data from PWT version 4. Versions 4 and 5.5 of
the PWT are very different.
To illustrate, G/Y (net of spending on defence and
education) (average G/Y in 1980-84) in Nepal is 0.0895 in PWT version 4 and 0.4449 in
PWT version 5.5. We opt to use data from Barro and Sala-i-Martin (1995) because their
data are derived from a later version of PWT.
Also at this point, it is important to point out some concerns regarding the manner
by which Gwartney et al. derive their Size variable. Gwartney et al. derive their
expected Size variable based on their hypothesis,
that the expected size of an economys trade sector is a function of

the following five variables: (1) geographic size, (2) population, (3) a
dummy variable indicating that the country is land-locked and
therefore restricted in its use of oceans for transport purposes, (4) a
dummy variable indicating that the country has potential trading
partners within 150 miles of its borders but less than 50 percent of its
population resides within this distance from the potential trading
partners, and (5) a dummy variable indicating that more than 50
percent of the countrys population was located within 150 miles of a
potential trading partner. The base for the last two variables is
therefore countries that do not have a potential trading partner within
150 miles.
Gwartney et al. (1996) p. 35.
However, when they run their regression to estimate the expected size of the
trade sector, they instead use the log of exports plus imports divided by GDP as the
dependent variable (p. 44), and conclude based on the R-squared of their regression
equation, that 50% of the variation in the size the trade sector across countries is due to
the variables they have identified. Note that Gwartney et al. do not provide an adjusted
72
Table 3.1
Government and Openness Indicators
Code
Variable
G/Y
Government general
consumption expenditures
as a share of GDP
This is the most commonly used measure of

government involvement in existing empirical
analyses.
Source: Table II-A

(Gwartney et al. 1996)
Assuming that not all government expenditures directly

affect productivity (Barro and Lee, 1994), and that
government operations are often less efficient than
private operations, one could expect that as the ratio of
government expenditures to GDP increase, lower
efficiency and economic growth will be likely. Thus,
the ratio of government expenditure to GDP is
expected to be negatively correlated with TFP and
growth.
Government consumption
as a percentage of GDP,
net of spending on defence
and education
Expenditures on defence and education are excluded

because these are not considered as consumption,
rather more like public investments that may have
direct effects on private-sector productivity, and
consequently, on private investments (Barro, 1991,
Barro and Sala-i-Martin (1995).
G/Y
(BS)
Source: Barro and Sala-iMartin (1995)
Interpretation
A negative correlation between G/Y (BS) and TFP and

growth is expected as with the previous (G/Y) variable.
GOE
Government-operated
enterprises (GOEs) as a
share of the economy
Source: Table II-B
Studies by the World Bank (e.g. World Bank 1988),

show that many GOEs, especially in the developing
countries, are inefficient and perform poorly (capital
and labour productivity are lower) compared to private
enterprises. Moreover, funding of GOEs (from already
scarce government resources), and consequently
protection of GOEs in terms of subsidies, favourable
tax treatment, etc. are not only a burden to taxpayers
(from whom government revenues come), but they also
stifle the competitiveness of private firms.
A country with a large (small) share of GOE in the
economy is given a low (high) rating, i.e., if GOEs
dominate in the economy (low rating), TFP is also low,
so the expected correlation is positive. A detailed
explanation of the GOE ratings is presented in
Appendix A.1.
continued over page
73
Code
Variable
Interpretation
Price
Price controls the extent

that businesses are free to
set their own prices
Government regulations such as price controls that

limit private businesses from engaging in trade at
prices that are mutually acceptable to both consumers
and producers could serve as a disincentive to
production. The presence of price controls also distorts
relative price signals and may prevent resources being
allocated to their efficient use.
Source: Table II-C

A country with little or no use of price controls is given

a high rating and TFP in that country is expected to be
high; thus a positive correlation is also expected. A
detailed explanation of the Price ratings is presented in
Appendix A.2.
Tradetax Taxes on trade as a
percentage of exports plus
imports
Tradetax = (tx+tm)/(X+M)
where,
tx = taxes on exports
tm = taxes on imports
X = total exports
M = total imports
Source: Table IV-A
The value of this Tradetax variable will be large, which

is indicative of a closed economy, if: (1) high taxes
on international trading prevail in the economy, and/or
(2) the economys trade volume (exports plus imports)
is low.
On the first point, taxes on international trade limit the
exchange between domestic residents and foreigners;
they increase the gap between what the seller receives
and what the buyers pay; and could reduce the volume
of international trade and consumer and producer
surplus.
High tax rates likely result in domestic residents and
foreigners being discouraged to trade; hence the
correlation between Tradetax and TFP is expected to
be negative.
On the second point, low trade volumes may imply
limited access to imported inputs (assumed to embody
new technology) and so could also hold back a
countrys capacity to specialise in research-intensive
production; hence a negative relationship between
Tradetax due to low trade volumes is hypothesised.
Black
Difference between the

official exchange rate and
the black market rate
Source: Table IV-B
Trade between countries would be facilitated if there

were no controls in currency exchange. The difference
between the official exchange rate and the black
market exchange rate implies that limits or controls on
currency exchange exist. Consequently, the larger the
gap between the official and black market exchange
rates, the lower the expected economic growth and
level of TFP, implying negative correlation.
continued over page
74
Code
Variable
Interpretation
Size
Actual size of trade sector

(ratio of exports plus
imports to GDP)
compared to the expected
size
This component proxies for other restrictions on trade,

e.g., quotas, monopoly grants, buy local schemes,
etc. (Gwartney et al., 1996) that hinder efficient trade
between domestic residents and foreigners. The
expected size of the trade sector in a particular country
is estimated using information on the countrys size,
population and location relative to potential trade
partners. The expected size is then compared with the
actual size of the countrys trade sector.
Source: Table IV-C

A smaller actual trade sector compared to the expected

size implies that there are sufficient restrictions on
trade prevailing in the economy that prevent the
country from achieving the expected trade sector size.
A large actual trade sector relative to the expected size
would imply that trade restrictions are minimal, so a
positive correlation with TFP and growth is expected.
R-squared value, nor the results of any diagnostic tests to check the reliability of their
results. Gwartney et al. then derive the expected size of the trade sector by plugging
the country characteristics for the variables into the regression equations (p. 45), and
the difference between the actual and the estimated size of the trade sector (from the
regression equation), is their Size variable, which they use to proxy for other restrictions
on trade. This Size variable however, does not accurately reflect other trade restrictions.
In fact, this variable may simply be a measure of errors associated with omitted
variables or mis-specification of the regression equation.
Despite these concerns
regarding the Gwartney et al. Size variable, we continue to use it in our analyses
because we would like to see how different measures of openness affect TFP, and
decide which one best measures openness.
The review of the empirical literature on the role of government in economic

growth points to the need for better measures of the extent of government involvement
in the economy (see Chapter 2, Section 2.3). In these empirical studies, aggregate
government expenditure has largely been used to measure government intervention.
However, this is a poor proxy for intervention because not all interventions require
spending. Besides, money spent on different types of intervention may also produce
different results. In this study, the extent of government involvement in the economy is
75
therefore measured not only in terms of government expenditures, but also in terms of
the extent of governments influence in setting prices in the economy and ownership of
industry.
Ideally, a tax variable, specifically a reliable measure of marginal tax rates,

would also be included in the analysis. As Helms (1985, p. 577) notes: it is not
meaningful to evaluate the effects of tax or expenditure changes in isolation: both the
sources and uses of funds must be considered. Kneller et al. (1999) and Bleaney et al.
(2001) also recognise the need for including both revenue and expenditure in the
analysis of fiscal effects on economic growth. Hence, a reasonable measure of marginal
tax rates prevailing in the economy is desirable. However, lack of reliable data on
marginal tax rates for a large cross-section of countries prevents its inclusion in this
chapter.39
3.3.2
Correlation analysis results
Table 3.2 presents the correlation results between the different TFP measures
(our improved TFP measure and Islams) and the different measures of government and
openness. Note that our TFP measures are for 1988, while Islams TFP measures are
the initial estimates for the period 1960-1985. For this reason, the period covered by the
government and openness variables used is adjusted to correspond to the measure of
TFP used.
Correlation coefficients are computed for the largest cross-section of
countries for which data are available. Preliminary inspection of the data suggests that
the data are not normally distributed.40
This being the case, the Spearman rank
correlation coefficient is used to analyse the correlation among the variables under
study.41
39
The issue surrounding the importance of a tax variable, particularly, a reliable measure of marginal tax
rates is examined in detail in Chapter 4.
40
This involved plotting a histogram for each of the variables against a normal curve centred on the
mean.
41
The Spearman rank correlation test does not make any assumption about the normality of data, unlike
other tests for correlation, e.g. the Pearson correlation test. Hence, in this case, it is the appropriate test to
use (McGhee, 1985).
76
Table 3.2
Variable
(expected sign)
Spearmans rank correlation coefficients for TFP, output,

government and openness measures
Own-TFP
(1)
Islam-TFP
(2)
Ygrowth
(3)
Y/cap
growth
(4)
Y/L growth
(5)
0.252**
(94)
0.214**
(86)
-0.212**
(87)
-0.059
(87)
-0.085
(86)
-0.682***
(90)
-0.759***
(86)
-0.081
(87)
-0.305***
(87)
-0.266**
(87)
GOE
(+)
0.210**
(94)
0.243**
(86)
0.050
(87)
0.072
(87)
0.113
(87)
Price
(+)
0.480***
(72)
0.589***
(65)
-0.003
(67)
0.340***
(67)
0.316**
(67)
Tradetax
(-)
-0.672***
(88)
-0.732***
(85)
0.034
(86)
-0.340***
(86)
-0.284***
(86)
Black
(-)
-0.469***
(94)
-0.505***
(86)
-0.119
(87)
-0.427***
(87)
-0.402***
(87)
0.118
(94)
0.110
(86)
0.206*
(87)
0.260**
(87)
0.196*
(87)
G/Y
(-)
G/Y (BS)
(-)
Size
(+)
Notes:
Numbers in parentheses are sample size; ***, **, and * indicate significant at 1%, 5% and 10%,
respectively, based on a two-tailed test.
Comparing the correlation results using Islams TFP measure with those
obtained when our estimates of TFP are used shows that the same government and
openness variables are significantly correlated with either TFP measure.
Note,
however, that the correlation coefficients when Islams TFP measure is used are greater
compared to the correlation coefficients obtained when our TFP measure is used.
The last three columns of Table 3.2 also present the correlation results between
the different measures of economic growth (growth in output (Ygrowth), growth in
output per capita (Y/cap growth) and growth in output per worker (Y/L growth)) and the
government and openness measures. Only G/Y and Size appear to be correlated with
Ygrowth. Only G/Y and GOE do not appear to correlate with either Y/cap growth or Y/L
77
growth. It is interesting to note that while G/Y is positively associated with TFP
(regardless of whatever measure of TFP is used),42 G/Y is negatively associated with

output growth. One reason for this difference may be due to the way the TFP variable
is constructed. That is, in deriving the TFP measure, the influence of other factors such
as labour and physical capital are already explicitly controlled and accounted for, while
the output measure is the sum of the contributions of all factors of production: labour,
physical and human capital and TFP. It is very likely that government and openness
variables relate differently to the different factors of production that constitute overall
output, which then affects the correlation results.
A closer look at the correlation results using our TFP measures indicates that all
the government measures are significantly correlated with TFP, and only one of the
openness measures, Size, appears to be uncorrelated with TFP.
Correlation results between the two government expenditure measures and TFP
are mixed, i.e., the results suggest that government expenditure as a percentage of GDP
(G/Y), has a positive correlation with TFP, yet the correlation coefficient between
government expenditures, net of spending on defence and education (G/Y (BS)) is
negative, as originally anticipated. Two plausible explanations for this particular result
are presented here.
First, if the two measures of government expenditures are derived from the same
source, i.e. if Gwartney et al. and Barro and Sala-i-Martin (BS) used the same
underlying data sets to calculate their respective measures of government expenditure,
then this result suggests that there are likely to be other items of government
consumption that should be more properly seen as investment, but are rarely
distinguished. For instance, Easterly and Rebelo (1993), Kocherlakota and Yi (1996),
Kneller et al. (1999) and Bleaney et al. (2001) point out the need to disaggregate
government expenditures as different types of expenditure may have a different effect
42
The positive association between G/Y and TFP is manifested throughout our analyses and will be
addressed further in the next sections.
78
on growth.43
The preceding result therefore suggests that pure government
consumption spending, as in G/Y (BS), is negatively associated with TFP. Recall that
Barro and Sala-i-Martin (1995) consider defence and education spending as investments
rather than consumption, but this is not to say that defence and education spending are
the only forms of government investment.
Total government expenditure, i.e.,
inclusive of all forms of government investments is positively associated with growth,

which could imply that the positive effects of government investments outweigh the
negative effects of pure government consumption.
However, Gwartney et al. and Barro and Sala-i-Martin derive their respective
measures of government expenditures from two very different datasets and this has
implications for the results of the analysis. Gwartney et al. derive their data from the
World Tables of the World Bank (and in some cases, from the International Monetary
Funds International Financial Statistics), which are measured in local currencies.

Barro and Sala-i-Martins data are from Summers and Hestons (1991) Penn World
Table version 5 (PWT5), which are measured in international prices. While it appears
that the use of local currencies should not have been a problem considering that the
government expenditure data are measured as ratios to GDP anyway, converting data to
international prices is sometimes problematic, as argued by Knowles (2001). This is
because the conversion to international prices can substantially change the
governments share in output for some countries. For example, if we compare the
Gwartney et al. and the PWT5 G/Y average measure for 1980-84, we see that the
Gwartney et al. G/Y for Nepal is 9.8% while it is 47.5% in PWT5. By contrast, the
Gwartney et al. G/Y for Canada is 20.1% while it is 11.7% in PWT5.44 A similar pattern
can be observed for most of the developing and developed countries, i.e. for the
developing countries, the PWT5 G/Y measure is greater than that of Gwartney et al. but
for most of the developed countries, the PWT5 measure is lower than that of Gwartney
et al.
43
44
Refer to Chapter 2, Section 2.3 for details of these studies.
We use average data for the period 1980-84 for ease of comparison. We also compare the Gwartney et
al. measure to that of the PWT5 instead of BS, to compare like with like, i.e. both the Gwartney et al. and
PWT5 data are for G/Y inclusive of spending on defence and education.
79
This pattern is observed in the data because the price of labour relative to capital
in the developing countries is generally lower than the world average relative price, but
the world average relative price is generally lower than the price of labour relative to
capital in developed countries. Also, because the government sector tends to be more
labour-intensive, conversion to international prices (as in the PWT) would increase
the relative price of governments output in the developing countries and decrease it in
the developed countries. As seen in the case of Nepal (a developing country), and
Canada (a developed country), conversion to international prices raises governments
share in output in the developing countries, while decreasing governments share in
output in the developed countries.
This is because the conversion of data to
international prices, amounts to setting the relative price in each country equal to the
international mean of relative prices, which means that, relative prices in each country
are determined by the world average, and not the relative prices that prevail in each
country individually (Knowles, 2001).
At this point, we are not trying to say that nothing worthwhile can be gleaned
from the use of international prices in this context, only that, for the particular purposes
of this study, doing so may be inappropriate. Whether it is appropriate to use local or
international prices may depend on what the data are trying to measure.
As already mentioned, we are mainly concerned with determining the size of

government in the economy, i.e, we want to be able to measure the quantity of resources
being used by the government.
appropriate.
Hence, in this case, the use of local prices is
To illustrate this idea, consider the cost of financing government
expenditure. If G/Y in local prices is 10%, government needs to raise 10% of output, Y,
in taxes. However, moving to international prices may increase G/Y to 50%, but this
does not necessarily mean that government needs to raise 50% of Y in taxes. Hence, if
we are interested in the cost of government spending, local prices are appropriate,
international prices are not.
80
Hence, in terms of the government expenditure measures, one reason why Barro
and Sala-i-Martins measure of G/Y (net of spending on defence and education) may
exhibit a negative correlation with TFP is because the G/Y (BS) data are inappropriate
(because of the problems caused by conversion to international prices).
For the
purposes of this study, we are more concerned with determining the size of government
in an economy, given the prices that prevail in that economy, and not particularly on
determining what the governments share in output is, if the world average relative
prices are used. Again, as Knowles (2001) notes, the conversion to international
prices distorts the data by imposing a set of international relative prices on each
country.45
Based on the aforementioned reasons, the use of data converted to international

prices for this particular purpose is therefore not desirable (i.e. the BS G/Y data) and
data on the share of government spending measured in local currencies (from Gwartney
et al. (1990)) are preferred and used in subsequent analyses.
Arguing that the conversion to international prices could cause the G/Y (BS) data
to exhibit a negative correlation with TFP tells only half the story since, following the
same argument above, the TFP measures are likely to be inappropriate as well. This is
because the TFP measures used in this analysis, i.e., our TFP measure (based on Hall
and Jones, 1999) and Islam (1985), were also derived from the PWT data measured in
international prices. To illustrate, we have said earlier that the price of labour relative to
capital (PL/PK) in the developing countries is generally lower than the world average
relative price. Put in a different way, this means that the price of capital relative to
labour (PK/PL) in developing countries is higher than the world average relative price.
This would make capital-intensive goods in the developing countries more expensive
compared to the world average price of capital-intensive goods, and so converting to
international prices would artificially decrease the relative price of capital-intensive
goods in developing countries. Consequently, the share of capital in GDP (K/Y) would
also tend to decrease in developing countries because of conversion to international
45
Dowrick and Quiggin (1997) also show that Summers and Hestons (1991) use of a common set of
relative prices understates the differences among the levels of income per capita across countries.
81
prices, and a decrease in K/Y leads to an increase in the TFP measure of these
developing countries. The opposite pattern occurs for the developed countries. Because
data converted to international prices, in particular the G/Y, K/Y and TFP data, are
subject to measurement biases, at least in terms of what we actually want to measure in
this study, the true or underlying association between these variables may therefore be
hidden.
Before proceeding further, it is also worth examining how the correlation

between the different government and openness variables is affected when an
alternative measure of TFP, one that is derived from data that are not influenced by
conversion to international prices, is used in the correlation analysis. The alternative
TFP data is derived as before, i.e., from equation (3.5) (without human capital in the
production function), but using K/Y data from Collins and Bosworth (1996). The K/Y
data from Collins and Bosworth are measured in local prices, which as we have
discussed earlier, are more appropriate to use in this study. The correlation between the
first TFP measure we derived and this alternative measure was also examined and the
two variables exhibit a high degree of correlation with each other.46
Table 3.3 compares the results of the correlation analysis between the new TFP
measure using K/Y data from Collins and Bosworth (1996) and each of the government
and openness variables with the correlation results using our first TFP measure in Table
3.2. Since the previous results (Table 3.2) are based on a different set of countries for
which the alternative TFP measure is available, correlations between our first TFP
measure and the different government and openness variables are again checked, this
time using the same set of countries for which the second TFP measure is available.
46
The Spearman rank correlation coefficient between the two TFP measures is 0.978.
82
Table 3.3
Spearmans rank correlation coefficients for the

alternative TFP measures
(1)
1st TFP measure

(using same set of
countries as the
alternative TFP)
(2)
2nd TFP measure

(using K/Y data from
Collins and
Bosworth, 1996)
(3)
0.252**
(94)
0.380***
(81)
0.404***
(81)
-0.682***
(90)
-0.630***
(79)
-0.668***
(79)
GOE
(+)
0.210**
(94)
0.248**
(81)
0.257**
(81)
Price
(+)
0.480***
(72)
0.468***
(66)
0.503***
(66)
Tradetax
(-)
-0.672***
(88)
-0.666***
(81)
-0.726***
(81)
Black
(-)
-0.469***
(94)
-0.579***
(79)
-0.607***
(79)
0.118
(94)
0.140
(81)
0.156
(81)
Variable
(expected sign)
G/Y
(-)
G/Y (BS)
(-)
Size
(+)
1st TFP measure

(from Table 3.2)
Notes:
Numbers in parentheses are sample size; ***, **, and * indicate significant at 1%, 5% and 10%,
respectively, based on a two-tailed test.
The results are presented in column (2) of Table 3.3. From this table we see that
the correlation results are very similar regardless of which TFP measure is used. This
implies that, even though it seems more appropriate to construct a measure of TFP with
K/Y measured in local prices, this does not seem to make much of a difference to the
correlation results. Nonetheless, in the subsequent analyses, we use the second TFP
measure, i.e., the one that is derived using K/Y in local prices, which is more appropriate
to use.
The correlation among the government and openness variables was also
examined. The results are presented in Table 3.4. Of the correlations among the
government variables, the results suggest that Price is significantly correlated with
83
G/Y(BS) and GOE. On the other hand, G/Y does not seem to be correlated with any of
the other government variables, G/Y(BS) is only correlated with Price, and GOE is also
only correlated with Price. It is also interesting to note the insignificant correlation
between the two measures of government expenditure, G/Y and G/Y (BS), which may
be due to the measurement differences pointed out earlier. The results we obtain from
this correlation exercise are basically consistent regardless of whether data for a single
time period (1988, top figure in Table 3.4), or data averaged over a specific period
(1975-1990, bottom figure in Table 3.4) are used.
From a theoretical perspective the different ways in which government

intervention has been measured appear to capture different aspects of governments
role, with some likely to affect TFP and/or output growth positively, while other
measures could have negative effects. So, as mentioned previously, including them all
in the regression equation could be worthwhile as each variable could provide additional
information in the TFP equations to be estimated. Unlike the different government
variables which measure different aspects of government intervention, each openness
variable we have is essentially trying to measure only one and the same thing: the
economys degree of openness.
From our correlation analyses we find that both
Tradetax and Black are significantly negatively correlated with TFP and output growth,
and Size is significantly positively correlated with output growth. These results suggest
that a more open economy, whether measured in terms of lower taxes on international
trade (as a percentage of the total trade sector), fewer exchange rate distortions, or a
larger trade sector, is expected to have a positive association with TFP and/or output
growth. Also, the statistically significant correlations among the different openness
measures likewise suggest that these variables measure or reflect the degree of openness
similarly.
Hence, even though there are different ways to measure the degree of
openness, we expect that effect of openness on TFP and/or output growth will be
approximately the same regardless of which particular measure is used, and so it is not
necessary to include all of the openness variables in the same regression equation.
What is of concern is choosing which one best measures the degree of openness of the
economy for our purposes. Having only one openness indicator would help avoid
losing degrees of freedom in the regressions.
84
Table 3.4
G/Y
Spearmans rank correlation coefficients for government and

openness variables
G/Y
G/Y (BS)a
GOE
Priceb
Tradetax
Black
Size
1.00
1.00
-0.017
-0.045
-0.146
-0.156
0.096
0.051
-0.370***
-0.387***
-0.245**
-0.234**
0.257***
0.349***
1.00
1.00
-0.135
-0.098
-0.373*** 0.550***
-0.374*** 0.489***
0.334***
0.320***
-0.177*
-0.112
0.478***
0.450***
-0.271***
-0.272***
-0.014
-0.066
G/Y (BS)b
GOE
1.00
1.00
Priceb
1.00
1.00
Tradetax
Black
-0.098
-0.123
-0.504*** -0.560*** 0.313***

-0.541*** -0.567*** 0.270**
1.00
1.00
0.572*** -0.393***
0.573*** -0.336***
1.00
1.00
Size
-0.445***
-0.406***
1.00
1.00
Notes
G/Y (BS) are averaged data for the period 1980 to 84.
b
Price data are for 1989.
***, **, and * indicate significance at 1% , 5% and 10% level based on a two-tailed test.
a
3.4
TFP and its determinants
3.4.1
Framework
The TFP measure used in the next analyses is different from the original HJ
(1999) TFP levels in two main ways. First, the K/Y data used are taken from Collins
and Bosworth (1996), which are considered preferable because these are measured in
local currencies and are therefore not subject to any measurement bias that is likely to
be caused by conversion to international prices. Second, unlike Hall and Jones (1999)
who use a production function with human capital-augmented labour to compute TFP
levels, we exclude human capital in our underlying production function. As mentioned
in the previous chapter, we subscribe to the findings of Benhabib and Spiegel (1994)
85
that cast doubt on the traditional role given to human capital in the development process
as a separate factor of production. They find evidence that the channel through which
human capital affects growth runs through productivity, rather directly as an input on a
production function. Except for the K/Y data, which we get from Collins and Bosworth
(1996), all the data we use to calculate our TFP levels are from Hall and Jones (1999).47
We also assume = 1/3 as in Hall and Jones (1999). The TFP levels calculated for this,
and the following estimations are presented in Appendix B.
Although we lack a sound theoretical framework on which to base our empirical

study, we opt to analyse TFP and its determinants using an approach similar to that of
Thomas and Wang (1993), Coe et al. (1997) and Miller and Upadhyay (2000),48 where
TFP is regressed on other variables that are perceived to affect TFP growth. In our
study, these other variables are: government, openness and human capital. Based on the
previous correlation results, we regress TFP levels on different government variables
and only one openness variable at a time. Acknowledging that human capital may
contribute to a larger effect of openness on TFP (see for instance Miller and Upadhyay,
2000), we also include interaction terms between human capital and openness. The
general TFP equation used in the estimation is presented below:
ln TFPi = a1 + a g Gig + a o Oio + a oh Oio hi + a h hi + i
(3.6)
where,
lnTFPi is the log of total factor productivity for country i,

Gig
is government measure g (in levels), for country i,
Oio
is openness measure o (in levels), for country i, and o is either Black,

Tradetax or Size,
Oiohi
are interaction terms between openness and human capital (in levels),
and,
47
Refer to Section 3.2.1 for a brief description of the Hall and Jones (1999) data.
48
Refer to Chapter 2, Section 2.2.3 for details on these papers.
86
hi
is a human capital variable from HJ (1999) (in levels), for country i.
Again, the TFP measure used in the succeeding analyses is that derived
following HJs (1999) levels accounting methodology, but without human capital in the
underlying production function and using K/Y data from Collins and Bosworth (1996),
hereafter TFP1.49
The government and openness variables included in estimating
equation (3.6) are as specified in Table 3.1, with the exception of the G/Y (BS), due to
some measurement concerns about the data discussed in the previous section. The
human capital data are taken from Hall and Jones (1999).50
We use the Shazam
Econometrics Computer Program, Version 7.0 (White, 1993) in estimating the equations
in this study.
3.4.2
Regression results
Table 3.5 presents the results of the TFP estimations. Columns (1), (2) and (3)
present results that use Black, Tradetax and Size as the measure of openness,
respectively in the TFP regressions. The signs in the brackets below each variable name
show the expected effect of the variable on TFP. For the interaction terms (human
capital-openness), the sign of the coefficient should be consistent with the sign of the
coefficient of the openness measure.
That is, the sign of the coefficients of the
interaction terms and the particular openness variable used should be the same to be able
to conclude that returns to openness are enhanced by human capital (Harrison, 1996).
Table 3.5 presents several interesting results based on the standard t-ratios (i.e.,
the numbers in parentheses).51 First, in terms of the significance of the government
variables, G/Y appears to be a consistent and highly significant positive influence on
49
In order that the TFP levels are comparable across countries, the Y/L data used in the underlying
production function are taken from Hall and Jones (1999), measured in international prices.
50
Hall and Jones (1999) human capital data are Mincer-based estimates of human capital-augmented
labour. See Section 3.2.1 for details.
51
Heteroskedastic-consistent t-ratios are presented in the same table, but discussion regarding these and
other results of the diagnostic tests is left to a separate section later so that emphasis can be given to the
need for diagnostics.
87
TFP, regardless of which openness measure was used in the regression. Contrary to
what was hypothesised, we find that a higher level of government spending is a positive
influence on TFP. While this result contradicts our initial hypothesis, this is nonetheless
not a surprising result as other empirical studies e.g., Helms (1985) Easterly and Rebelo
(1993), Kocherlakota and Yi (1996), Kneller et al. (1999), Bleaney et al. (2001) have
shown that growth is positively associated with some types of government
expenditures.52 The coefficients of the GOE and Price variables are also significant and
positive. Recall that GOE and Price are measured as ratings and a higher rating reflects
less government intervention. Positive coefficients on the GOE and Price variables
therefore mean that less government intervention in terms of government ownership of
enterprises and less Price control in the economy promote higher TFP.
Second, in terms of openness variables, only Black and Tradetax appear to be a

significant influence on TFP.
We find that these measures of openness have the
expected negative sign, which implies that a less open economy is a negative influence
on TFP levels.
Third, the human capital variable appears to be significant only in the TFP
regressions using Black as the proxy for openness. This result suggests no robust
significant relationship between human capital and TFP, contrary to what was
hypothesised. Nonetheless, this result is consistent with the findings of Miller and
Upadhyay (2000) who also find that human capital has no significant effect on TFP that
excludes human capital in the first stage production, and other researchers, e.g. Islam
(1995) and Benhabib and Spiegel (1994), although Islam and Benhabib and Spiegel
examine the effect of human capital on growth instead of on TFP. With respect to the
interaction terms in the regression equations, again, none of the interaction terms
between human capital and the openness variables appears to be statistically significant.
52
See Chapter 2 for a review of other empirical studies that have found a positive relationship between
government expenditures and growth.
88
Table 3.5
TFP estimation results
Variable
(expected sign)
(1)
(2)
(3)
G/Y
(-)
0.0683
(3.538) ***
[5.193] ***
0.0543
(2.733) ***
[3.402] ***
0.0740
(3.849) ***
[5.558] ***
GOE
(+)
0.1137
(2.234) **
[2.662] ***
0.0828
(1.662) *
[1.675] *
0.1010
(1.971) **
[2.376] **
Price
(+)
0.0863
(2.087) **
[2.333] **
0.0662
(1.595) *
[1.922] *
0.0982
(2.299) **
[2.844] ***
Black
(-)
-0.0044
(-1.614) *
[-2.585] **
Tradetax
(-)
-0.1219
(-1.397) *
[-1.649] *
Size
(+)
-0.0049
(-0.6174)
[-0.9471]
Human
(+)
0.1092
(0.6011)
[0.8146]
0.0381
(0.1556)
[0.2156]
0.1546
(0.8170)
[0.9745]
Human-Open#
(+ or -)
0.0018
(1.603) *
[2.759] ***
0.0314
(0.6678)
[0.7704]
0.0030
(0.6780)
[0.9480]
Constant
6.6797
(12.87) ***
[16.09] ***
7.4887
(11.63) ***
[12.82] ***
6.4690
(12.69) ***
[13.95]***
64
64
64
0.4222
0.4621
0.4006
0.3613
0.4055
0.3375
F (5, 58)
8.444***
9.970***
7.779***
Notes:
Numbers in parentheses are t-ratios and numbers in square brackets are heteroskedasticconsistent t-ratios; ***, ** and * indicate significance at 1%, 5% and 10% levels respectively
(based on a one-sided test); Open# is either Black, Tradetax or Size.
To summarise, the results of the TFP estimations in Table 3.5 show that,
contrary to what was hypothesised, if we look at the significance of the coefficient
estimates based on standard t-ratios: (1) G/Y is a positive influence on TFP, (2) the
degree of openness of the economy (measured in terms of either Black or Tradetax) does
89
influence TFP; (3) human capital is not a significant influence on TFP, and (4) the
interactions between human capital and openness are not significant influences on TFP
either. Only our hypotheses on the negative influence of GOE and Price on TFP are
supported by the results of the estimations.
3.5
Robustness and Diagnostic tests
3.5.1
Robustness tests
(a)
using TFP calculated from a production function with human capital-augmented

labour and K/Y data from Collins and Bosworth (1996), hereafter TFP2
We also tried to see how robust our results are if we use instead a TFP measure
which is derived from a production function with human capital-augmented labour (as in
the original Hall and Jones approach, but instead, we use the K/Y data from Collins and
Bosworth, 1996). Re-estimating equation (3.6), using TFP2, we find that the same
government and openness variables appear significant in this new set of regressions (see
Appendix C).53 The major difference between the results in Table 3.5 and Appendix C
is with regard to the human capital variable, which now appears to be a significant
negative influence on TFP2 (derived from a production function with human capital as a
factor of production). While we expected the relationship between human capital and
TFP to be positive, the negative result here is not entirely surprising, and this finding is
again consistent with the findings of Miller and Upadhyay (2000), when they use TFP
that includes a human capital variable in the first-stage production function.
Next, we also attempt to see what happens if we exclude human capital and the
interaction terms in the regression equation, i.e., we estimate:
ln TFPi = a1 + a g Gig + a o Oio + i
(3.7)
53
Diagnostic tests for this new set of regression equations were also carried out and are reported in
Appendix C.1.
90
where,
lnTFPi is the log of total factor productivity for country i (calculated from a
production function with human capital-augmented labour),
Gig
is government measure g (in levels), for country i,
Oio
is openness measure o (in levels), for country i, and o is either Black,

Tradetax, or Size.
The results of estimating the above equation are presented in Appendix D, and
show similarities to those in Table 3.5. That is, the same explanatory variables appear to
be significant in the regressions in Appendix D. The results in Table 3.5 and Appendix
D confirm that G/Y still appears to have a significant and positive influence on TFP. We
also find that less government intervention in terms of ownership and control of
enterprises, and less use of price controls are positive influences on TFP.
(b)
using K/Y data from Hall and Jones (1999) in calculating TFP without human
capital in the underlying production function, hereafter TFP3, and TFP from a
production function with human capital-augmented labour, hereafter TFP4
Next, we attempt to find out whether there are any substantial differences in the
results of Table 3.5 and Appendix C if we use the K/Y data from Hall and Jones in
computing TFP levels (with and without human capital in the underlying production
function). The results with TFP3 as the dependent variable are presented in Appendix E
and the results with TFP4 as the dependent variable, and excluding human capital as an
explanatory variable in the TFP regression, are presented in Appendix F.
In brief, we find that the significance of the government variables based on

standard t-ratios, varies across the estimations. GOE and Price appear to be significant
positive influences on TFP (except when the Tradetax measure of openness is included
as one of the regressors). G/Y appears to be a significant positive influence only when
the dependent variable is TFP3 (see Appendix E); even then, this still contradicts our
initial hypothesis. The human capital variable is not found to be significant in any of the
91
estimations in Appendix E. In terms of the openness variables, Size is found to have a
significant influence on TFP3 and the interaction term between human capital and Size
is also found to be a significant influence on TFP3.54
Tradetax is found to have a
significant negative influence only on TFP4 (which is derived from a production

function with human capital-augmented labour).
This exercise shows that the significance of the variables being studied,
especially for the government variables, appears somewhat sensitive to the particular
TFP measure used, even though the different TFP measures are highly correlated.55
3.5.2
Diagnostic tests
We have seen how our results in Table 3.5 compare with other sets of results that
use different TFP measures. But how do our results compare with the results of other
empirical studies on TFP and its determinants? Our initial results in Table 3.5 contradict
some of the findings of earlier studies. For instance, although some of our findings are
consistent with Miller and Upadhyay (2000) (see Sections 3.4.2 and 3.5.1 (a)), Miller
and Upadhyay find that human capital and the interaction between human capital and
openness are significant and positive influences on TFP;56 however, we find no
significant relationship between these variables and TFP in our estimations in Table
3.5.57 So, whose results are more reliable? Many of the empirical studies that analyse
TFP and its determinants (e.g., Collins and Bosworth, 1996; Thomas and Wang, 1993;
54
We refrain from drawing much attention to the significance of the interaction term between human
capital and Size because the sign of the coefficient is contrary to what was hypothesised.
55
For example, the correlation between TFP1 and TFP2 is 0.959 (without and with human capital
respectively in deriving TFP, and both TFP derived using K/Y data from Collins and Bosworth);
correlation between TFP1 and TFP3 is 0.978 (both TFP derived without human capital in the underlying
production function, but TFP1 uses K/Y data from Collins and Bosworth (1996) and TFP2 uses K/Y data
from Hall and Jones (1999)).
56
Miller and Upadhyay (2000) find that human capital and the interaction term between human capital
and openness are positive influences on TFP that is calculated from a production function without human
capital, but find that these variables are negative and significant (at the 20% level) in the regressions with
TFP calculated from a production function with human capital.
57
We only find a significant relationship between an interaction term (Size and human capital) and TFP
in one regression in Appendix E.
92
Coe et al. 1997; and Miller and Upadhyay, 2000) do not report results of diagnostic tests
on their regression equations.58 This makes the reliability of the results in previous
studies questionable. To examine the reliability of our results, and to determine whether
or not there are any problems with the TFP regression equations we estimated, a series
of diagnostic tests is performed.
The most obvious diagnostic tests to run are for normality of the error terms,
heteroskedasticity (a common problem with using cross-sectional data), and for whether
the model is mis-specified. Table 3.6 reports the results of these diagnostic tests (Hetconsistent t-ratios for the coefficients of the variables are presented in square brackets in
Table 3.5).59
The results of the Jarque-Bera test for the normality of the error terms (based on
a Lagrange Multiplier test) show that the OLS residuals in all of the TFP estimations are
non-normal.60 These results cast doubt on the statistical significance (or lack of it) of the
explanatory variables in our regression equations. Also, if the error distributions are
non-normal and likely skewed, then the reliability of the goodness of the least-squares fit
is also questionable, because the residuals of the regression equation may have been
affected by outliers and/or influential observations.61
Further diagnostic tests show that the null hypothesis of homoskedasticity is

rejected by half of the tests in each regression, implying a significant heteroskedasticity
problem in our regressions. In the regression with Black as the proxy for openness
(column 1 of Table 3.6), this heteroskedasticity problem is not as significant. Of the
three tests that reject the null of homoskedasticity, only one is significant at the 5%
58
Collins and Bosworth (1996) report the adjusted R2 only; Thomas and Wang (1993) report only the
adjusted R2 and some F-tests; Coe et al. (1997) report the R2, the adj R2, the SEE (standard error of
estimation) and some F-tests; and, Miller and Upadhyay (2001) report only the adjusted R2 and SEE.
59
Diagnostic test on the other regression equations are reported in Appendices 2A to 5A.
60
Problems of heteroskedasticity and mis-specification of the model are also possibly dependent on nonnormality.
61
The problem of outliers and/or influential observations is discussed in the following section.
93
Table 3.6
Null hypothesis
(1)
(2)
(3)
9.1290 **
(2)
11.8715 ***
(2)
6.8777 **
(2)
2.601
(1)
2.819 *
(1)
2.376
(1)
6.401
(6)
15.228 **
(6)
12.464 *
(6)
5.175 **
(1)
5.487 **
(1)
4.830 **
(1)
6.957
(6)
10.486
(6)
12.637 *
(6)
4.729 **
(1)
4.825 **
(1)
4.621 **
(1)
6.621
(6)
6.823
(6)
10.280
(6)
0.9562
(1,56)
0.4877
(2,55)
0.4079
(3,54)
6.4652 **
(1,56)
3.3743 **
(2,55)
2.7441 *
(3,54)
0.5045
(1,56)
0.4348
(2,55)
1.4339
(3,54)
Jarque-Bera test for normality of the error term

LM-test
Homoskedasticity
( i )2 on Yi
( i )2 on [Yi ]
( i )2 on ln[Yi ]
( i )2 on X
ln ( i ) on X
2
i on X
Correct specification
RESET (2)
RESET (3)
RESET (4)
Notes:
Numbers in parentheses are degrees of freedom. The Jarque-Bera and heteroskedasticity tests
are 2 distributed and the RESET tests are F-distributed under their respective null hypotheses.
***, ** and * rejects the null hypothesis at 1%, 5% and 10% levels, respectively.
level, with the two significant only at the 10% level. The problem of heteroskedasticity
is highlighted most in the regression that uses Tradetax as the proxy for openness
(column 2 of Table 3.6). Four of the six tests reject the null of homoskedasticity, and
three of these are significant at the 5% level.
In terms of the RESET tests for possible mis-specification of the model, the null
hypothesis that there is no mis-specification is rejected in all the tests only when the
Tradetax measure of openness was used in the TFP regressions. This suggests that, of
the three sets of TFP regressions, only the regression that includes the Tradetax measure
94
of openness as an explanatory variable is mis-specified, as far as the RESET test is
concerned.
The diagnostic tests on the regressions that use different TFP measures for the
dependent variable (Appendices 2A to 5A) also reveal problems with these regressions.
All the regressions suffer from a significant problem of having non-normally-distributed
errors. The problem of heteroskedasticity is also present in all but one of the regressions
(column 1 of Appendix E.1), and the problem of mis-specification is apparent, again
only in the regressions that use Tradetax as the measure of openness.
The results of the diagnostic tests show that our initial findings regarding the
significance of the government variables, and lack of significance of the openness,
human capital and interaction terms are not particularly reliable. How then do we
address the problems identified by the diagnostic tests?
In terms of heteroskedasticity, one way of dealing with this problem is by using

Whites (1980) heteroskedastic-consistent covariance matrix estimation technique. This
can be done using the HETCOV option under the OLS estimation in the SHAZAM
econometric package. This gives heteroskedastic-consistent (Het-consistent) t-ratios,
which are used to test the significance of the coefficients when heteroskedasticity is
present. The numbers in square brackets in Table 3.5 present the Het-consistent t-ratios
for each of the coefficient estimates. Looking at the significance of the coefficients of
the explanatory variables based on these Het-consistent t-ratios reveals that all the
government variables are statistically significant influences on TFP. G/Y still remains a
positive influence on TFP, and less government intervention in terms of ownership and
control of corporations and prices in the economy (higher GOE and Price ratings imply
less government intervention) are also positive influences on TFP. We now find that at
least one openness measure, Black, significantly affects TFP, and as hypothesised, we
find that the black market premium is a negative influence on TFP. We also find that
the coefficient of the interaction term between Black and human capital is significant
based on the Het-consistent t-ratio, but its sign is positive.
95
Although we have addressed the heteroskedasticity problem (up to a point), the
reliability of the significance of our coefficient estimates still remains flawed because of
the problem of non-normal errors in our regressions. Often the error terms are not
normally distributed because of the presence of outliers and/or influential observations.
An outlier is an observation that is located far or deviates notably from the other
observations in the sample, while an observation is considered influential if its residual
is large and/or is located far from the other observations in the sample (Sen and
Srivastava, 1990). Note that an observation can be influential even though its residual is
not large and an observation may be an outlier but not influential. There are several
methods by which outliers and influential observations can be detected.
For the
purposes of this study, the hat value, which is a measure of leverage in regressions, and
the RSTUDENT, or the Studentised residuals, which identify outlying observations are
used to determine whether there are any influential observations in the previously
estimated regression equations.
In a simple OLS regression, the hat value, hi, is a diagonal element of the hat
matrix, H = X(XX)-1X (where X is the n x k matrix of observations on a set of
explanatory variables). If hi is large then the ith observation is considered to have a
high leverage or substantial impact on the estimates. Sen and Srivastava (1990), suggest
hat values less than 0.2 to be small enough not to have any influential impact on the
estimates, and any hat value above 0.2 is considered influential. The same cut-off point
is adopted in this study.62
The RSTUDENT is defined as:
e*i =
62
ei
s (i ) 1 hi
(3.8)
Belsley, et al. (1980) suggest an alternative cut-off point, hi = 2(k+1)/n, where k is excluding the
constant. Using this formula gives a cut-off point hi 0.22 which is roughly the same as the cut-off point
we have used.
96
where ei is a residual from an ordinary OLS regression, s(i) is the residual standard
deviation obtained from an OLS regression that deletes the ith case; and hi is a diagonal
element of the hat matrix. RSTUDENT is a useful tool to show whether or not an
observation fit in the model. As a general rule of thumb, an observation i is influential if
the absolute value of the ith RSTUDENT is greater than two (Sen and Srivastava, 1990).
The computed hat values and RSTUDENTs show that there are a number of
influential observations in the estimated regression equations in Table 3.5.63 For the
regression that uses Black as the measure of openness (column 1), five influential
observations were found: Kenya and Malawi, based on the RSTUDENT statistics, and
Iran, Israel and Sierra Leone, based on the hat values. The regression that uses Tradetax
as the proxy for openness (column 2) had the most number of influential observations. In
this regression, there are ten influential observations: Kenya, Malawi, Nigeria and
Tanzania based on the RSTUDENTs; and Bangladesh, Ghana, India, Iran, Israel and the
U.K. based on the hat values. For the regression that uses Size as the proxy for openness,
there are eight influential observations: Kenya, Malawi and Nigeria based on the
RSTUDENTs; and Belgium, Israel, Jordan, Malaysia and Singapore based on the hat
values.
Having identified these influential observations however does not mean that they
should just be discarded.
The presence of influential observations could point to
possible errors in the data, parameter heterogeneity, or the need for model respecification, and should therefore be examined more closely. Where the data for the
influential observations are not in error (i.e., we trust the reliability of our data sources)
and/or the re-specification of the model proves to be more problematic than deleting the
influential observations, then, deleting these observations may be appropriate.64 Temple
(1999a) also finds trimming down the sample size to exclude influential observations
63
See Appendix G for the list of countries included in each specific regression equation with the
influential observations indicated by a *.
64
Lorgelly and Owen (1999), based on Beckman and Cook (1983) and Fiebig (1992), provide a
discussion of alternative ways of dealing with influential observations and the advantages and
disadvantages of each.
97
and outliers necessary, to be able to distinguish the patterns present in the majority of the
data.65
3.5.3
TFP estimations without the influential observations
The issue of influential observations is addressed by deleting the influential

observations identified in each of the regressions. The results of the regressions without
the influential variables are presented in Table 3.7. Before discussing the results in
Table 3.7, we perform the diagnostic tests again to determine whether there remain any
problems with this new set of regressions.
The results of the diagnostic tests are
presented in Table 3.8.
The Jarque-Bera test for the normality of residuals shows that the error terms in
this new set of TFP regressions are normally-distributed and hence imply that the use of
conventional tests of statistical significance of these coefficients is more reliable
compared to those in the initial results in Table 3.5.
The diagnostic tests for
heteroskedasticity show that the null hypothesis of homoskedasticity is rejected by about

half of the tests in the TFP regressions with either Black or Size included as an
explanatory variable (columns 1 and 3 of Table 3.8), and rejected by more than half of
the tests in the TFP regression with the Tradetax variable. These results show that the
problem with heteroskedasticity still remains in this new set of regressions. Further still,
similar to the results of the previous diagnostic tests, we find that the mis-specification
problem is present only in the regressions that use the Tradetax measure of openness.
To account for the influence of heteroskedasticity on the standard errors, Het-consistent
t-ratios are again reported in square brackets in Table 3.7.
65
However, Temples (1999a) preferred method of trimming down sample size is different from ours, as
he prefers to use least-trimmed squares. Still, he argues that excluding influential observations and
outliers is a desirable strategy.
98
Table 3.7
TFP estimation results without the influential observations
Variable
(expected sign)
(1)
(2)
(3)
G/Y
(-)
0.0681
(3.419) ***
[4.385] ***
0.0496
(2.756) ***
[3.418] ***
0.0836
(4.013) ***
[5.267] ***
GOE
(+)
0.1186
(2.671) ***
[2.899] ***
0.0486
(1.089)
[1.163]
0.0979
(1.966) *
[2.188] **
Price
(+)
0.0438
(1.129)
[1.292] *
0.0322
(0.8558)
[1.042]
0.0855
(1.912) *
[2.365] **
Black
(-)
-0.0064
(-0.7305)
[-0.7813]
Tradetax
(-)
-0.1946
(-1.865) **
[-1.727] **
Size
(+)
0.0064
(0.6037)
[0.8075]
Human
(+)
0.0333
(0.2003)
[0.2503]
-0.1749
(-0.6980)
[-0.6509]
0.0946
(0.5667)
[0.6395]
Human-Open#
(+ or -)
0.0008
(0.1341)
[0.1608]
0.0536
(0.9604)
[0.7619]
-0.0035
(-0.5841)
[-0.7463]
Constant
7.1213
(14.09) ***
[17.39] ***
8.4680
(13.02) ***
[14.54] ***
6.6066
(14.06) ***
[14.44] ***
64
54
56
R2
0.5010
0.4848
0.4660
R2
0.4435
0.4190
0.4006
F (5, 58)
11.517***
10.894***
10.130***
Notes:
See Appendix G for a list of countries excluded in the estimations.
Numbers in parentheses are t-ratios; numbers in square brackets are Het-consistent t-ratios.
***, ** and * indicate significance at 1%, 5% and 10% levels respectively (based on a one-tailed
test).
99
Table 3.8

For regressions without influential observations
Null hypothesis
(1)
(2)
(3)
2.9935
(2)
3.1538
(2)
1.8481
(2)
5.828 **
(1)
5.913 **
(1)
5.726 **
(1)
7.833
(6)
11.089 *
(6)
10.605
(6)
5.183 **
(1)
5.208 **
(1)
5.153 **
(1)
12.518 *
(6)
12.946 **
(6)
13.118 **
(6)
6.817 **
(1)
6.684 **
(1)
6.950 ***
(1)
8.491
(6)
7.050
(6)
8.207
(6)
0.0115
(1,51)
0.1835
(2,50)
0.8721
(3,49)
3.2790 *
(1,46)
3.3988 **
(2,45)
2.5676 *
(3,44)
0.5045
(1,51)
0.4348
(2,50)
0.3514
(3,46)

LM-test
Homoskedasticity
( i )2 on Yi
( i )2 on [Yi ]
( i )2 on ln[Yi ]
( i )2 on X
ln ( i ) on X
2
i on X
RESET (2)
RESET (3)
RESET (4)
Notes:
Numbers in parentheses are degrees of freedom. The Jarque-Bera and heteroskedasticity tests
are 2 distributed and the RESET tests are F-distributed under their respective null hypotheses.
***, ** and * reject the null hypothesis at 1%, 5% and 10% levels, respectively.
The results of this new set of TFP regressions without the influential
observations identified earlier are considered more robust and reliable as the problems
identified in the previous diagnostic tests have been addressed. In terms of the overall
fit of the regression equation, as measured byR2, we find that this new set of
regressions fit better (i.e., have a higherR2) compared to the regressions in Table 3.5.
G/Y, contrary to our initial hypothesis, is determined to be a highly significant

positive influence on TFP, based on the Het-consistent and conventional t-ratios. As
discussed previously, while this result does not support our initial hypothesis, it is not
inconsistent with other studies (e.g., Helms, 1985; Easterly and Rebelo, 1993;
100
Kocherlakota and Yi, 1996; Kneller et al., 1999; Bleaney et al., 2001) that have found
that some types of government expenditures positively affect growth. The results in
Table 3.5 show that GOE and Price are also significant (although slightly less
significant than G/Y) influences on TFP. However, results of our new regressions based
on Het-consistent t-ratios, show that GOE and Price are significant only when the Black
or Size measure of openness is used in the regression. Only the Tradetax measure of
openness appears to be a significant negative influence on TFP. The human capital
variable and the interaction terms between human capital and openness were not
significant in any of the TFP regressions.
3.6
Summary and conclusions
This chapter attempts to determine empirically the factors that affect TFP levels
across countries. Motivated by earlier empirical studies that relate TFP to government
and openness variables, we estimate a simple TFP equation using different measures of
government intervention and the degree of openness of the economy, a human capital
variable, and interaction terms between openness and human capital.
We analyse the determinants of TFP based on a levels framework because levels

more accurately capture the long-run effects of the variables on TFP, compared to the
use of growth rates. The TFP levels we use in our analyses are different from Hall and
Jones (1999) TFP levels in two main ways. First, our TFP levels are derived by using
data on the ratio of physical capital to output (K/Y) that are not subject to any bias that
may be caused due to conversion to international prices. We therefore use K/Y data
from Collins and Bosworth (1996) that are measured in local currencies. We also
exclude human capital from the underlying production function used in our levels
accounting exercise. We maintain, following Nelson and Phelps (1966), that simply
including human capital as an additional factor of production would represent a
misspecification of the productive process. We also propose that any effect of human
capital on overall economic growth is actually transmitted via its effects on TFP, hence,
human capital should be kept out of the production function.
101
We include in our analyses different measures of government intervention:
government consumption expenditure as a share of output (G/Y), G/Y net of spending on
defence and education (G/Y (BS)) (at least in the initial correlation exercise), and
measures for the level of government intervention in the economy via ownership and
control of corporations and prices (GOE and Price). Many of the empirical studies that
relate government to growth or TFP primarily use government expenditures to proxy for
governments role in the economy, but this is not ideal since governments can intervene
without necessarily spending, e.g. through ownership and control of corporations or
through influencing prices in the economy. We examine the correlation between TFP
and the different government and openness measures to determine first and foremost
whether or not it would be meaningful to relate these variables to each other. We find
that, except for the GOE variable, TFP and the various government and openness
measures are significantly correlated. We also find that the correlation between TFP
and the two G/Y variables are contradictory. TFP and G/Y are positively correlated, but
TFP and G/Y (BS) (i.e., G/Y net of spending on defence and education) are negatively
correlated. We suggest that this contradictory result may be due in part to the way G/Y
(BS) has been derived from data measured in international prices.
Initial regression results suggest that government expenditure (G/Y) positively

influences TFP, contrary to our hypothesis that government expenditures are detrimental
to growth and TFP.
While this result contradicts our hypothesis of a negative
relationship between government spending and TFP, it is nonetheless not inconsistent

with a number of empirical studies. Less government intervention in the form of few
government-owned enterprises and few price controls in the economy appear to
positively affect TFP. Using either Black or Tradetax as a proxy for openness, we find
that a less open economy is a negative influence on TFP levels. Human capital and the
interactions between the human capital and openness variables also do not significantly
affect TFP levels. We check whether our results are robust if alternative measures of
TFP, e.g. TFP levels calculated from a production function with human capitalaugmented labour and TFP levels based on the original Hall and Jones data are used in
the regressions. We find that the significance of the explanatory variables differ once a
different TFP measure is used.
102
We also perform diagnostic tests on our regression equations, something that

many previous empirical studies seem to lack. We find that our initial regressions
suffer from problems of non-normally-distributed error terms and heteroskedasticity.
The problem of model mis-specification is apparent only in the regression that uses
Tradetax as the proxy for openness. We address the heteroskedasticity problem by
computing Whites (1980) heteroskedastic-consistent (Het-consistent) t-ratios for each
of the coefficient estimates. The problem of non-normal residuals is dealt with by
deleting the identified influential observations in the regressions.
Based on our new set of regression results, we still find that, government
spending is a highly significant positive influence on TFP.
GOE and Price are
significant except when the Tradetax measure of openness is used in the regression.
As mentioned previously, the diagnostic tests show that the problem regarding
model mis-specification is apparent only in the regressions that include Tradetax as one
of the explanatory variables. This suggests that the results that use Tradetax in the
regression equation should be considered with caution. Recall that one of the objectives
in this chapter is to compare different measures of openness to determine which one
best captures the effect of openness on TFP. The regression results and diagnostic tests
show that compared to the regressions that use either Black or Size as the proxy for
openness, the regression results with Tradetax as an explanatory variable appears most
problematic. That is, the problems of non-normal errors and heteroskedasticity are more
significant in this regression, and the problem of model mis-specification is present only
in this regression, even after deleting the influential observations.
Our regression results however show that neither Black nor Size is a significant
influence on TFP. This has two main implications: (i) either Black or Size is not an
appropriate measure of openness, and/or (ii) the degree of openness of the economy has
no significant effect on TFP. As noted in Section 3.3.1, we have concerns regarding the
way Gwartney et al. (1996) calculated the Size variable; hence we are not totally
convinced about the reliability of any result that includes Size in the regression
103
equations despite the necessary checks we have made to ensure that our results are
robust.
Of the three proxies for openness then, we find that Black may be the
appropriate and reliable measure to use, and based on this we find that, contrary to our
hypothesis, the degree of openness of the economy has no robust effect on TFP.
The human capital variable and the interaction terms between human capital and
openness were also determined to be not statistically significant in all of the TFP
regressions. Again, we are led to conclude that, contrary to our expectations, human
capital has no significant effect on the level of TFP. There is also no evidence to
support our hypothesis that openness, enhanced by human capital (as represented by our
interaction terms) affects TFP levels.
To conclude, this chapter shows that differences in TFP levels across countries
are substantially influenced by government intervention. Contrary to our hypothesis,
government intervention in terms of total government consumption expenditures is a
positive influence on TFP. The other forms of government intervention, e.g., ownership
and control of enterprises and price controls, are, as hypothesised, negative influences
on TFP, but these results are quite fragile depending on the other variables included the
regression equation. On the other hand, neither human capital, nor any interaction
between human capital and the degree of openness of an economy has any significant
influence on TFP levels. Openness, measured in terms of either Black or Tradetax
appears to be a significant influence on TFP.
These conclusions are drawn from a careful study of TFP levels across countries
using an alternative measure of TFP levels.
We also use different proxies for
government intervention and not only the usual government expenditures found in many
empirical studies of growth and government. Finally, diagnostic tests, which many
previous empirical works lack, are carried out to examine the reliability and robustness
of our results.
We note that in this cross-sectional analysis, we have not looked into the
possibility of reverse causation, i.e., that it is possible that the relationships run not only
104
from our hypothesised determinants of TFP to TFP, but also from TFP to these
variables. For instance, the positive G/Y, TFP link could be an example of causality
running from high TFP to high government consumption; i.e., rich/advanced
countries demanding more government services.
However, analysing issues of
causality is difficult in cross-sectional studies, hence is left to the time-series approach

in the next chapter.
105
~ Appendix A.1 ~
Explanation of GOE Ratingsa
Rating
Role of Government Enterprises in Country
10
There are very few government-operated enterprises and they produce less than 1
percent of the countrys total output.
There are very few government-operated enterprises other than power-generating

plants and those operating in industries where economies of scale generally
reduce the effectiveness of competition.
Government enterprises are generally present in power generating, transportation

(airlines, railroads, and bus lines), communications (television and radio stations,
telephone companies, and post offices) and the development of energy sources,
but private enterprises dominates other sectors of the economy.
There are a substantial number of government-operated enterprises in many

sectors of the economy, including the manufacturing sector. Most of the large
enterprises of the economy are operated by the government; private enterprises
are generally small. Employment and output in the government-operated
enterprises generally comprises between 10 and 20 percent of the total nonagricultural employment and output.
Numerous government enterprises of all sizes are present and they operate in
many sectors of the economy, including manufacturing and retail sales.
Employment and output in the government-operated enterprises generally
comprise between 20 and 30 percent of the total non-agricultural employment
and output.
The economy is dominated by government-operated enterprises. Employment

and output in the government-operated enterprises generally exceeds 30 percent
of the total non-agricultural employment and output.
Reproduced from Gwartney, Lawson and Block (1996), p. 265.
106
~ Appendix A.2 ~
Explanation of Price Ratingsa
Rating
General characteristics of Country
10
No price controls or marketing boards are present.
Except in industries (e.g., electric power generation) where economies of scale

may reduce the effectiveness of competition, prices are generally determined by
market forces.
Price controls are often applied in energy markets; marketing boards often
influence prices of agricultural products; controls are also present in a few other
areas, but most prices are determined by market forces.
Price controls are levied on energy, agricultural, and many stable products (e.g.,
food products, clothing and housing) that are widely purchased by households;
but most other prices are set by market forces.
Price controls apply to a significant number of products in both agricultural and

manufacturing industries.
There is widespread use of price controls throughout the economy.
Reproduced from Gwartney, Lawson and Block (1996), p. 268.
107
~ Appendix B ~
1988 TFP levels (in logs)
COUNTRY
Algeria
Argentina
Australia
Austria
Bangladesh
Belgium
Benin
Bolivia
Botswana
Brazil
Burundi
Cameroon
Canada
Cen. Afr. Rep.
Chad
Chile
Colombia
Congo
Costa Rica
Cyprus
Denmark
Dom. Rep
Ecuador
Egypt
El Salvador
Fiji
Finland
France
Gabon
Ghana
Greece
Guatemala
Haiti
Honduras
Hong Kong
Iceland
India
Indonesia
Iran
Ireland
TFP1
TFP2
TFP3
TFP4
8.591655
8.964577
9.700302
9.51456
8.106657
9.789576
n.a.
7.833966
n.a.
8.790889
n.a.
7.420997
9.951525
n.a.
n.a.
8.745923
8.723964
n.a.
8.676963
9.13429
9.557863
8.421316
8.455038
8.447566
8.204283
n.a.
9.548119
9.730443
n.a.
7.159251
9.169884
8.504084
7.256993
8.07439
n.a.
9.683203
7.583554
7.873663
8.675576
9.425662
8.271395
8.157897
8.607982
8.71192
7.842677
8.771376
n.a.
7.338166
n.a.
8.522779
n.a.
6.319837
9.782685
n.a.
n.a.
8.156393
8.303204
n.a.
7.824833
8.03585
9.003683
7.725736
7.808188
7.969186
7.389523
n.a.
8.757599
9.428323
n.a.
6.347521
8.822824
8.287004
6.779953
7.18388
n.a.
9.274503
7.081054
7.434143
7.734896
8.389782
8.88657
9.19963
9.76334
9.60785
8.92110
9.82294
7.69611
8.20476
7.56030
9.01693
7.21332
7.83171
9.96170
7.15856
7.61397
8.70081
8.89417
8.30373
8.81560
9.24213
9.57538
8.64048
8.57121
9.13958
8.61511
8.80853
9.55115
9.73632
7.90709
7.73515
9.24426
8.89084
7.70922
8.28548
9.82625
9.68681
7.91409
7.90979
8.82306
9.42414
8.56631
8.39295
8.67103
8.80522
8.65712
8.80475
7.60231
7.64048
7.06450
8.54926
6.94521
7.53289
8.86054
6.98973
7.26530
7.91736
8.30464
7.88296
8.14527
8.39000
8.47693
8.08630
7.87563
8.49273
8.13673
7.99376
8.50983
8.94580
7.60497
7.30367
8.43249
8.54378
7.49214
7.80844
8.93574
8.75793
7.50539
7.40729
8.38354
8.48346
continued over page
108
~ Appendix B ~
(continued)
COUNTRY
Israel
Italy
Ivory Coast
Jamaica
Japan
Jordan
Kenya
Madagascar
Malawi
Malaysia
Mali
Malta
Mauritius
Mexico
Morocco
Netherlands
New Zealand
Nicaragua
Niger
Nigeria
Norway
Pakistan
Panama
Paraguay
Peru
Philippines
Portugal
Rwanda
Senegal
Sierra Leone
Singapore
Somalia
South Africa
South Korea
Spain
Sri Lanka
Sweden
Switzerland
Syria
Taiwan
TFP1
TFP2
TFP3
TFP4
9.568489
9.738336
7.571038
7.627916
9.358717
9.149269
6.992128
6.951697
6.39795
8.642951
6.616941
9.167176
8.623746
9.090895
8.328935
9.653261
9.585963
7.957876
n.a.
6.712317
9.495063
8.121367
8.429383
8.204318
8.448976
7.884563
8.851509
7.215746
7.52269
7.374618
9.347817
n.a.
8.46279
9.134633
9.564634
8.120046
9.702357
9.722545
n.a.
9.386733
8.802209
9.345926
7.018878
6.656636
8.791407
8.735209
6.459948
6.605977
5.72459
8.533071
5.788041
8.571586
8.045326
8.446775
7.350175
9.146741
9.513603
7.761666
n.a.
6.455037
8.726763
7.514667
7.712593
7.417838
7.935756
7.772003
8.531249
6.985266
6.93114
7.068048
8.715867
n.a.
7.76721
8.460263
8.526034
7.105926
9.167697
8.883545
n.a.
8.741653
9.66106
9.78171
7.99538
7.91946
9.37986
9.39248
7.44031
8.02897
6.89294
8.70079
7.30825
9.35816
9.17160
9.32838
8.82998
9.75006
9.58144
8.06281
6.85294
7.05970
9.57013
8.51438
8.51496
8.46375
8.65043
8.11241
9.05949
7.69084
8.12949
8.51447
9.49315
7.38387
8.68128
9.20753
9.62413
8.56927
9.75624
9.71658
9.40807
9.41292
8.62517
9.01542
7.60297
7.36730
8.40857
8.82517
7.02625
7.49678
6.54722
8.02743
7.19837
8.52926
8.57601
8.74995
8.18586
8.77130
8.36672
7.55630
6.78058
6.86349
8.46829
8.25711
7.74667
7.85704
7.93364
7.32593
8.54628
7.57829
7.80923
8.28399
8.90159
7.07730
8.04932
8.28269
8.92855
7.89489
8.71764
8.70245
8.87341
8.57392
continued over page
109
~ Appendix B ~
(continued)
COUNTRY
Tanzania
Thailand
Togo
Trinidad/Tobago
Tunisia
Turkey
Uganda
United Kingdom
Uruguay
U.S.A.
Venezuela
Zaire
Zambia
Zimbabwe
TFP1
TFP2
TFP3
TFP4
6.391796
8.192621
n.a.
9.132705
8.331179
8.382214
6.727262
9.687524
8.840749
9.978785
9.116163
6.759777
6.644403
7.171954
6.106376
7.404121
n.a.
8.691845
8.073899
7.398014
5.943812
8.489804
8.166379
9.678625
8.544813
6.407357
n.a.
n.a.
7.00351
8.47459
6.92905
9.37275
8.77790
8.57971
7.58763
9.82179
9.06333
10.02462
9.32630
7.29759
6.85581
7.40526
6.69665
7.82950
6.64363
8.58425
8.44558
8.13885
7.33035
8.83759
8.27988
8.82690
8.65192
6.99743
6.28446
7.05284
110
~ Appendix C ~

with human capital-augmented labour,
with human capital and interaction terms as regressors in the TFP equation
Variable
(expected sign)
(1)
(2)
(3)
G/Y
(-)
0.0685
(3.533) ***
[5.240] ***
0.0540
(2.707) ***
[3.390] ***
0.0737
(3.825) ***
[5.554] ***
GOE
(+)
0.1133
(2.217) **
[2.656] **
0.0820
(1.641) *
[1.668] *
0.1011
(1.968) **
[2.384] **
Price
(+)
0.0899
(2.164) **
[2.398] **
0.0686
(1.648) *
[1.974] *
0.1005
(2.346) **
[2.870] ***
Black
(-)
-0.0042
(-1.537) *
[-2.547] **
Tradetax
(-)
-0.1139
(-1.302) *
[-1.538] *
Size
(+)
-0.0004
(-0.5565)
[-0.8288]
Human
(+)
-0.3951
(-2.166) **
[-2.944] ***
-0.4526
(-1.841) **
[-2.509] **
-0.3502
(-1.846) **
[-2.211] **
Human-Open#
(+ or -)
0.0018
(1.530)
[2.726] ***
0.0268
(0.6543)
[0.2064]
0.0028
(0.6309)
[0.8549]
Constant
7.0184
(13.47) ***
[16.97] ***
7.8192
(12.11) ***
[13.46] ***
6.8195
(13.35) ***
[14.83] ***
64
64
64
R2
0.4231
0.4640
0.4035
R2
0.3624
0.4076
0.3407
Notes:
(based on a one-tailed test).
Open# is either Black, Tradetax or Size.
111
~ Appendix C.1 ~
(for the regressions in Appendix C)
Null hypothesis
(1)
(2)
(3)
9.876 ***
(2)
12.177 ***
(2)
7.428 **
(2)
2.010
(1)
2.163
(1)
1.849
(1)
6.083
(6)
11.473 *
(6)
11.738 *
(6)
4.246 **
(1)
4.504 **
(1)
3.943 **
(1)
6.254
(6)
12.098 *
(6)
11.135 *
(6)
3.470 *
(1)
3.580 *
(1)
3.345 *
(1)
6.245
(6)
7.698
(6)
9.603
(6)
0.825
(1,56)
0.542
(2,55)
0.379
(3,54)
7.657 ***
(1,56)
3.808 **
(2,55)
3.1851 **
(3,54)
1.948
(1,56)
0.962
(2,55)
0.952
(3,54)

LM-test
Homoskedasticity
( i )2 on Yi
( i )2 on [Yi ]
( i )2 on ln[Yi ]
( i )2 on X
ln ( i ) on X
2
i on X
RESET (2)
RESET (3)
RESET (4)
Notes:
Numbers in parentheses are degrees of freedom.
The Jarque-Bera and heteroskedasticity tests are 2 distributed and the RESET tests are Fdistributed under their respective null hypotheses.
112
~ Appendix D ~

with human capital-augmented labour,
without human capital and interaction terms as regressors in the TFP
equation
Variable
(expected sign)
(1)
(2)
(3)
G/Y
(-)
0.0705
(3.604) ***
[5.158] ***
0.0518
(2.550) **
[3.149] ***
0.0710
(3.656) ***
[5.234] ***
GOE
(+)
0.0946
(1.825) **
[2.267] **
0.0719
(1.422) *
[1.521] *
0.0958
(1.853) **
[2.302] **
Price
(+)
0.1078
(2.585) **
[2.921] ***
0.0757
(1.795) *
[2.127] **
0.1072
(2.484) **
[2.989] ***
Black
(-)
-0.0001
(-0.248)
[-0.0448]
Tradetax
(-)
-0.0605
(-2.302) **
[-2.381] **
Size
(+)
0.0004
(0.1539)
[0.2499]
Constant
6.2178
(15.54) ***
[16.92] ***
6.9919
(13.84) ***
[13.25] ***
6.1969
(16.06) ***
[17.97] ***
64
64
64
R2
0.3681
0.4196
0.3677
R2
0.3253
0.3802
0.3248
Notes:
113
~ Appendix D.1 ~
(for the regressions in Appendix D)
Null hypothesis
(1)
(2)
(3)
5.294 **
(2)
8.278 **
(2)
4.872 **
(2)
3.153 *
(1)
3.242 *
(1)
3.054 *
(1)
6.077
(4)
10.597 **
(4)
8.852 **
(4)
2.959 *
(1)
3.183 *
(1)
2.711 *
(1)
4.355
(4)
7.687
(4)
7.097
(4)
3.479 *
(1)
3.555 *
(1)
3.392 *
(1)
5.221
(4)
8.843 *
(4)
7.550
(4)
1.439
(1,58)
0.720
(2,57)
1.368
(3,56)
5.464 **
(1,58)
2.713
(2,57)
3.337 **
(3,56)
1.492
(1,58)
0.756
(2,57)
1.164
(3,56)

LM-test
Homoskedasticity
( i )2 on Yi
( i )2 on [Yi ]
( i )2 on ln[Yi ]
( i )2 on X
ln ( i ) on X
2
i on X
RESET (2)
RESET (3)
RESET (4)
Notes:
114
~ Appendix E ~
TFP estimation results, using all data from Hall and Jones (1999) for
calculating TFP (1999) from a production function
without human capital-augmented labour
Variable
(expected sign)
(1)
(2)
(3)
G/Y
(-)
0.0457
(2.471) **
[2.880] ***
0.0316
(1.752) *
[1.860] *
0.0498
(2.813) ***
[3.518] ***
GOE
(+)
0.0681
(1.397) *
[1.516] *
0.0550
(1.181)
[1.114]
0.0617
(1.312) *
[1.465] *
Price
(+)
0.0899
(2.281) **
[2.456] **
0.0395
(1.011)
[1.256]
0.0882
(2.224) **
[2.719] ***
Black
(-)
-0.0003
(-0.1108)
[-0.1812]
Tradetax
(-)
-0.0793
(-0.9696)
[-0.9999]
Size
(+)
-0.0129
(-1.778) *
[-2.291] **
Human
(+)
0.1541
(0.8903)
[1.125]
0.1269
(0.5430)
[0.7394]
0.0441
(0.2543)
[0.2970]
Human-Open#
(+ or -)
0.0002
(0.1345)
[0.2344]
0.0083
(0.1905)
[0.2064]
0.0067
(1.712) *
[2.284] **
Constant
7.2179
(14.58) ***
[16.72] ***
8.0589
(13.35) ***
[14.10] ***
7.4027
(15.95) ***
[17.10] ***
70
68
70
R2
0.2701
0.3499
0.3047
R2
0.2006
0.2859
0.2384
Notes:
(based on a one-tailed test); Open# is either Black, Tradetax or Size.
115
~ Appendix E.1 ~
(for the regressions in Appendix E)
Null hypothesis
(1)
(2)
(3)
9.0251 **
(2)
10.9317 ***
(2)
8.4259 **
(2)
1.740
(1)
1.751
(1)
1.725
(1)
5.420
(6)
5.040
(6)
7.894
(6)
5.450 **
(1)
5.705 **
(1)
5.168 **
(1)
7.695
(6)
5.459
(6)
10.053
(6)
3.099 *
(1)
3.172 *
(1)
3.021 *
(1)
6.388
(6)
8.914
(6)
9.554
(6)
0.7115
(1,62)
0.4806
(2,61)
0.7330
(3,60)
9.9972 ***
(1,60)
5.0083 ***
(2,59)
4.1301 ***
(3,58)
0.9775
(1,62)
0.6598
(2,61)
0.7601
(3,60)

LM-test
Homoskedasticity
( i )2 on Yi
( i )2 on [Yi ]
( i )2 on ln[Yi ]
( i )2 on X
ln ( i ) on X
2
i on X
RESET (2)
RESET (3)
RESET (4)
Notes:
116
~ Appendix F ~
TFP estimation results, using all data from Hall and Jones (1999) for
calculating TFP (1999) from a production function
with human capital-augmented labour
Variable
(expected sign)
(1)
(2)
(3)
G/Y
(-)
0.0164
(1.083)
[1.102]
0.0056
(0.3611)
[0.3513]
0.0186
(1.234)
[1.366]
GOE
(+)
0.0554
(1.361) *
[1.412] *
0.0490
(1.221)
[1.095]
0.0551
(1.371) *
[1.406] *
Price
(+)
0.0695
(2.112) **
[2.295] **
0.0334
(0.9742)
[1.157]
0.0786
(2.332) **
[2.471] **
Black
(-)
0.0000
(0.1004)
[0.1110]
Tradetax
(-)
-0.0470
(-2.272) **
[-2.200] **
Size
(+)
-0.0018
(-0.9878)
[-1.044]
Constant
7.3695
(23.14) ***
[23.03] ***
7.9358
(20.38) ***
[17.54] ***
7.3295
(24.01) ***
[24.69] ***
72
70
72
R2
0.1732
0.2267
0.1850
R2
0.1239
0.1791
0.1363
Notes:
117
~ Appendix F.1 ~
(for regressions in Appendix F)
Null hypothesis
(1)
(2)
(3)
5.9703 *
(2)
6.5384 **
(2)
4.8018 *
(2)
5.638 **
(1)
5.614 **
(1)
5.656 **
(1)
9.839 **
(4)
9.115 *
(4)
13.164 **
(4)
9.839 ***
(1)
10.204 ***
(1)
9.435 ***
(1)
13.842 **
(4)
7.728
(4)
15.063 ***
(4)
7.586 ***
(1)
7.580 ***
(1)
7.586 ***
(1)
13.260 **
(4)
13.773 ***
(4)
16.774 ***
(4)
0.0561
(1,66)
0.0366
(2,65)
0.1766
(3,64)
4.3829 **
(1,64)
2.1675
(2,63)
1.9472
(3,62)
0.0304
(1,66)
0.2764
(2,65)
0.9115
(3,64)

LM-test
Homoskedasticity
( i )2 on Yi
( i )2 on [Yi ]
( i )2 on ln[Yi ]
( i )2 on X
ln ( i ) on X
2
i on X
RESET (2)
RESET (3)
RESET (4)
Notes:
118
~ Appendix G ~
List of countries included in the main TFP regressions (Table 3.5)
(1)
with Black as a regressor
(2)
with Tradetax as a regressor
(3)
with Size as a regressor
Argentina
Australia
Austria
Bangladesh
Belgium
Malawi *
Malaysia
Malta
Mexico
Morocco
Argentina
Australia
Austria
Bangladesh *
Belgium
Malawi *
Malaysia
Malta
Mexico
Morocco
Argentina
Australia
Austria
Bangladesh
Belgium *
Malawi *
Malaysia *
Malta
Mexico
Morocco
Bolivia
Brazil
Cameroon
Canada
Chile
Netherlands
New zealand
Nigeria
Norway
Panama
Bolivia
Brazil
Cameroon
Canada
Chile
Netherlands
New Zealand
Nigeria *
Norway
Panama
Bolivia
Brazil
Cameroon
Canada
Chile
Netherlands
New zealand
Nigeria *
Norway
Panama
Colombia
Cyprus
Denmark
Dom. Rep
Ecuador
Paraguay
Peru
Philippines
Portugal
Senegal
Colombia
Cyprus
Denmark
Dom. Rep
Ecuador
Paraguay
Peru
Philippines
Portugal
Senegal
Colombia
Cyprus
Denmark
Dom. Rep
Ecuador
Paraguay
Peru
Philippines
Portugal
Senegal
Egypt
Finland
France
Ghana
Greece
Sierra Leone *
Singapore
South Africa
South Korea
Spain
Egypt
Finland
France
Ghana *
Greece
Sierra Leone
Singapore
South Africa
South Korea
Spain
Egypt
Finland
France
Ghana
Greece
Sierra Leone
Singapore *
South Africa
South Korea
Spain
Guatemala
Haiti
India
Indonesia
Sweden
Switzerland
Taiwan
Tanzania
Guatemala
Haiti
India *
Indonesia
Sweden
Switzerland
Taiwan
Tanzania *
Guatemala
Haiti
India
Indonesia
Sweden
Switzerland
Taiwan
Tanzania
Iran *
Ireland
Israel *
Italy
Thailand
Trin/Tob
Tunisia
Turkey
Iran *
Ireland
Israel *
Italy
Thailand
Trin/Tob
Tunisia
Turkey
Iran
Ireland
Israel *
Italy
Thailand
Trin/Tob
Tunisia
Turkey
Jamaica
Japan
Jordan
Kenya *
UK
Uruguay
US
Venezuela
Jamaica
Japan
Jordan
Kenya *
UK *
Uruguay
US
Venezuela
Jamaica
Japan
Jordan *
Kenya
UK
Uruguay
US
Venezuela
Notes:
* indicates that the country is influential and is therefore excluded in the corresponding TFP
regressions in Table 3.7.
119
~ CHAPTER FOUR ~
Testing for Cointegration and Causality for the HPAEs
4.1
Introduction
One of the aims of this research is to determine how the economic performance
of the High-Performing Asian economies (HPAEs) differs from the rest of the world,
particularly with respect to the importance of TFP in the HPAEs growth experience. In
this chapter we continue to re-examine the links among TFP, openness, government and
human capital using time-series data for the HPAEs, focusing particularly on the causal
relationships among these variables.
We pursue the analysis of the effects of human capital and openness on TFP
levels, even though results in the previous chapter suggest that human capital and
openness (measured in terms of the black market exchange rate premium (Black), taxes
imposed on international trade (Tradetax), or the size of the actual trade sector relative
to the expected size (Size)), are not significant determinants of TFP in a cross-country
study.
As previous studies have shown (e.g. Harrison, 1996), the significance of
variables being studied can differ substantially depending on the particular type of
dataset used. Also, as Temple (1999b) points out, the association between human
capital and growth is sometimes hidden in cross-country studies (as in Benhabib and
Spiegel, 1994 and Pritchett, 1997) due to a number of unrepresentative observations,
i.e., influential observations and/or outliers.
Time series analysis also has some distinct advantages over pure cross-section
analysis.66 For instance, as pointed out in the previous chapter (see Section 3.2.2), the
conversion of some variables into a common currency, a requirement in cross-sectional
studies, is sometimes problematic. Briefly, conversion to international prices (as in
the PWT data), distorts the data by imposing a set of international relative prices on each
country (Knowles, 2001). In time-series studies, data need not be converted to any
66
Humphries (1999) provides a detailed discussion of the advantages of time-series over cross-section
analysis.
120
other currency, as all come from one country.
Also, in cross-section studies,
particularly for those that study developed and developing countries together, data need
to have been compiled in a consistent fashion; otherwise, results will be erroneous.
With time-series studies, again, this problem is avoided because the data used in the
analysis come from one country.
Miller (1996) also points out that, in cross-country regressions, it is implicitly

assumed that countries have similar structural characteristics (e.g. in terms of
production technologies and institutional patterns). As such, political, economic, social
and other structural differences are therefore seen to have no effect on the growth
process or, if they do, their effects are randomly distributed with zero mean. However,
if these country-specific characteristics do condition the growth process in some nonrandom way, then cross-country studies may be seriously flawed. There have been
attempts to address this concern in the empirical literature. For example, oil-exporting
countries have often been excluded from cross-country studies, primarily because their
major source of income comes from exploiting their natural resources, which makes
them very different from most other countries. Other studies attempt to control for
differences by including regional dummy variables in their analyses. Still others choose
to examine the growth process within individual countries over time.
In this study we do not mean to imply that there is nothing to be gained from
doing cross-country studies, only that it may be useful to turn to time-series data in
order to determine the importance of some country-specific effects.
In this chapter, we also hypothesise that government has an influence not only
on TFP, but on the degree of openness in an economy. Islam (1998) points out that
previous export-led growth (ELG) studies have produced conflicting results because of
the bivariate nature of the models examined in these empirical works. Islam notes that
governments, through fiscal and monetary policies (e.g., maintaining a stable exchange
rate, various tax incentives and credit guarantees for firms in the export sector, etc.),
have influenced the export-growth relationship.
Results of his study of 15 Asian
121
countries67 over the period 1967-1991 indicated that export expansion causes growth in
ten of the fifteen countries studied, and that this relationship is fairly stable, largely
because of the role of government (proxied by total government expenditure).
However, we disagree with Islams use of total government expenditure to proxy
for the fiscal and monetary policies that he assumes to be the mechanisms by which
government has an impact on growth. To illustrate, credit guarantees for firms in the
export sector cannot be adequately proxied by government spending, because being a
guarantor does not necessarily involve spending. Also, changes in the black market
premium68 are a better gauge of the stability of the foreign exchange rate than
government spending; and marginal tax rates can better proxy for tax (dis)incentives.
Moreover, as discussed more extensively later in this chapter (also in Chapter 2, Section
2.3.3), different types of government expenditures can affect the economy in different
ways. Nonetheless, Islam (1998) remains an important contribution to the analysis of
the ELG hypothesis by analysing other key variables that are perceived to influence
the export-growth relationship within multivariate cointegration and causality
frameworks.
Our study offers several contributions to the empirical literature on the

determinants of TFP in the HPAEs. First, instead of testing the more common exportled-growth (ELG) hypothesis, where only exports and GDP growth have been tested for
possible cointegration and causality,69 we test what we refer to as the openness-ledgrowth (OLG) via TFP hypothesis. That is, we hypothesise that openness influences
economic growth, but essentially through its effect on TFP. Second, we build on
Islams hypothesis that other factors affect the manner in which openness influences
growth, by including various proxies for the different roles of government and human
capital in our analyses. Third, we conduct our analyses in the context of TFP levels and
not output growth and hypothesise stable, long-run cointegrating as well as causal
67
Japan, Hong Kong, South Korea, Singapore, Indonesia, Malaysia, Philippines, Thailand, Fiji (AsiaPacific), Papua New Guinea, India, Nepal, Pakistan, Sri Lanka and Bangladesh.
68
The black market premium is the difference between the official exchange rate and the back market
exchange rate.
69
See Giles and Williams (1999) for a survey of the empirical literature on ELG.
122
relationships between TFP and its perceived determinants (openness, government and
human capital), and among the determinants too. Finally, we analyse the hypothesised
cointegrating and causal relationships within multivariate cointegration and causality
frameworks following Johansens (1988) cointegration procedure and error correction
model tests for causality.
Section 4.2 describes the different government and openness data used in the
analysis. Section 4.3 shows how TFP levels and growth rates for the HPAEs are
calculated and how these compare with previous studies results. Section 4.4 describes
the Granger causality concept and shows how it is related to the concept of
cointegration and the error correction model (ECM) representation.
Section 4.5
examines the statistical properties of the data through standard unit root tests. Section
4.6 describes the Johansens multivariate cointegration procedure and results of the
estimation of the error correction representation of the models. Section 4.7 tests for
Granger causality in the HPAEs; and Section 4.8 presents a summary and conclusions
on the determinants of TFP levels in the HPAEs.
4.2
Data
Data for the different government and openness indicators and GDP were taken
from various issues of the International Monetary Funds (IMF) International Financial
Statistics (IFS) and Government Finance Statistics (GFS) yearbooks.
The GFS
yearbooks however do not report data for Hong Kong and Taiwan; hence our analysis is
limited to the other five HPAEs only: Indonesia, Malaysia, Singapore, South Korea and
Thailand from the 1970s to the 1990s. The human capital data, as well as the raw data
required in calculating the TFP levels, for the five HPAEs were taken from a later
version of the data from Collins and Bosworth (1996).
As discussed in Chapter 2, Section 2.3, total government expenditure appears to

be a poor proxy for the actual role of government in the economy. This is because
different types of government expenditures can affect the economy in different ways;
hence, there is a need to disaggregate government expenditure data. For example,
Helms (1985) in a study of the effect of (US) state and local taxes on economic growth,
123
found that taxes spent on government transfer payments are detrimental to growth,
while taxes spent on education, transportation and communication enhance economic
growth. Similarly, Kneller et al. (1999) and Bleaney et al. (2001) also show that
productive government expenditure in the OECD countries enhances growth, while
non-productive government expenditure does not.70
These studies have also
emphasised the importance of analysing the effects of both government revenue (taxes)
and expenditures on economic growth, and not just either one in isolation.71 For these
reasons, it is desirable that different proxies for the role of government, other than just
total government expenditure (or total consumption), be included in the analyses.
Annual time-series data for the various government measures used in the
previous chapter (i.e., the extent of government ownership of enterprises in the
economy (GOE) and the extent of the use of price controls (Price)) cannot be obtained,
so analysis in this chapter is limited to the use of government expenditure data
disaggregated
in terms of non-complementary (Noncomp) expenditures and
complementary (Comp) expenditures, and government revenue data.72 Government

non-complementary expenditure is defined as expenditures that may not directly affect
the productivity of the private sector (i.e., general public services, defence, and social
security and welfare) that may or may not be positively associated with TFP, if it is
determined that the effects are non-zero to begin with. On the other hand, government
complementary expenditures are those that are in some way conducive to private
investment (i.e., health, education and economic services, including expenditures on
70
Kneller et al. (1999) and Bleaney et al. (2001) define productive government expenditures as those
expenditures with substantial (physical or human) capital component (i.e., general public expenditures,
defence, education, health, housing and transportation and communication expenditures), while for nonproductive expenditures, the main component is social security expenditures.
71
We note, however, that Helms (1985), Kneller et al. (1999) and Bleaney et al. (2001) have not dealt
specifically with causality issues, hence it is possible that the identified effects of government
expenditures are actually caused by growth.
72
Kaufmann et al. (1999) propose new aggregate measures of governance, defined as the traditions
and institutions by which authority in a country is exercised. While their paper could make a substantial
contribution to empirical analyses of the effects of government on economic growth, we are unable to
make use of their data in this analysis because their governance data are cross-sectional. We are also
unable to use Kaufmann et al.s data in our cross-sectional analysis (Chapter 3) because they are available
only for the 1997-1998 period.
124
research, roads, other transport and communication services), and hence may be
positively associated with TFP.
There is no set rule regarding how to classify the different types of government
expenditures. Indeed, as Kneller et al. (1999 p.173) point out, the classification of
particular expenditures into productive or non-productive expenditures can be
debatable. In this study, the decision as to which expenditures should fall under the
Noncomp classification, and which should fall under the Comp, is primarily based on
our hypotheses regarding the effects of the different components of government
expenditures on TFP. For instance, we hypothesise that expenditures on health and
education promote a healthier and more educated workforce and expenditures on
economic services lead to better infrastructure and telecommunication facilities, which
help facilitate private transactions and assist in raising productivity levels. Hence,
government expenditures in these areas are classified under Comp.
Kneller et al. (1999) and Bleaney et al. (2001) classify economic services under
non-productive expenditures, arguing that expenditures on economic services in many
less developed countries (LDCs) are often dominated by assorted government subsidies
that may be distortionary in practice. Although the use of subsidies has been common
in the HPAEs (Tan, 1992; World Bank, 1993a; Tipton, 1998), it remains uncertain
whether economic subsidies in general are necessarily distortionary and therefore
should be classified as non-productive expenditure. Take for instance, subsidised
credit to priority sectors and sometimes to specific firms. To some extent, these have
been effective in allocating resources to high-yielding investments (World Bank,
1993a). The development of the export sectors of the HPAEs, due in part to the
governments policy of providing cheap credit, aided the accelerated adoption and
mastery of international best practice technologies. Should subsidised credit then be
considered non-productive expenditure?
Also, agricultural subsidies and other
expenditure on extension services have substantially aided the development of the

agricultural sector and even allowed food supplies to increase substantially faster than
population (Tipton, 1998). Are these non-productive expenditures as well? Although
some argue that such expenditures could also lead to inefficiency and, if left to the
125
private sector more efficient prices may be obtained, still, these expenditures may be
less efficient, but not non-productive. They are complementary expenditures that do
not necessarily prohibit market and exchanges, and hence are classified as such, i.e.,
Comp in this study.
On the other hand, the effects of other government expenditures (i.e., general
public services, defence, social security and welfare) on private sector productivity
levels are more ambiguous, i.e., they may or may not enhance productivity levels, and
hence are classified together under Noncomp.
Government revenues in this study are essentially the taxes received by the
central government. For our purposes, non-tax revenues are excluded from government
revenues, on the assumption that non-tax revenues do not affect private-sector
investment incentives. Taxes used to finance government spending have generally been
viewed as distortionary, discouraging private investment initiatives and therefore
detrimental to TFP. However, depending on what government revenues are used for,
these may be positively associated with economic growth (see for instance Helms, 1985,
Kneller et al., 1999 and Bleaney et al., 2001) and/or TFP. We disaggregate total
government tax revenue into: domestic tax revenues as a percentage of GDP, Domtax,
(including taxes on income, profit, payroll and manpower, property, social security
contributions, and taxes on domestic goods and services) and international tax revenues,
Tradetax, which are essentially taxes imposed on international trade as a percentage of
the total trade sector. We disaggregate government revenue in this manner, instead of
into distortionary and non-distortionary taxation as in Kneller et al. (1999) and Bleaney
et al. (2001), because we will be using Tradetax as our measure of openness. In this
study we also aim to determine the role of government, not only in promoting TFP but
also in affecting the degree of openness of an economy. As detailed in Chapter 3 Table
3.1, we hypothesise that Tradetax, although we recognise that this is a less than perfect
proxy for openness, will have a negative effect on TFP.
For Domtax, if a non-zero
effect of Domtax is determined, we hypothesise that the effect on TFP can be either
positive or negative. This is because Domtax is a mixture of income and consumption
taxes which Kneller et al. (1999) argue can be quite different in their effects on growth.
126
We are unable to further disaggregate Domtax into distortionary or non-distortionary
revenue in order to avoid losing valuable degrees of freedom.
In Chapter 3 we noted that the marginal tax rate is the more relevant variable to
study instead of tax revenue as a proportion of GDP (average tax); we explain this point
in more detail here. People are less concerned with the proportion of total income paid
in taxes (average tax) than with the amount of tax paid on the last dollar earned
(marginal tax). Hence the marginal tax rate is an important determining factor in
deciding whether or not do extra work. For example, if an extra hour of labour leads
to extra income being taxed at a higher rate, then the disincentive to do that extra work
is greater. Since there is often more than one marginal tax rate present in an economy,
the choice of the particular marginal tax rate to use in an empirical analysis is less
straightforward.
Nonetheless, several attempts have been made to compute the
marginal tax rates relevant to studies of economic growth.
Gwartney et al. (1996) gather data on the top marginal tax rates and the income
threshold at which they apply for a wide range of countries. These countries are then
given a zero to ten rating. A rating of ten indicates that the country has a low marginal
tax rate and the income threshold for which that rate applies is high, while a rating of
zero implies that the country imposes a high marginal tax rate and the income threshold
for which that rate applies is quite low. Data are obtained from various issues of Price
Waterhouses Individual Taxes: A Worldwide Summary. The main problem with the
Gwartney et al. measure is that, although it takes into account the income threshold at
which the top marginal tax rate applies, it is heavily dominated by what the top rate is.
Furthermore, Gwartney et al. measure the top marginal tax rate only, and so a country
may have a high marginal tax rate and consequently receive a low rating, but this high
tax rate may only affect a small percentage of the population. If only a small fraction of
the population receives income that is subject to high taxation, then this may not
accurately reflect the distortionary effects of taxation postulated (e.g. disincentives to do
extra work). The Gwartney et al. measure also considers individual marginal tax rates
only, and this implicitly assumes that other entities are affected similarly. In fact, the
marginal tax rate that affects the economic decision-making of an individual may be
127
quite different from the level that would affect, say, a private corporation in the
economy. Gwartney et al. therefore ignore the effects of taxes on the corporate level by
looking at individual taxes and income thresholds only.
Seater (1985) and Koester and Kormendi (1989) also ignore this important
qualification when they calculate their respective marginal tax rate measures. Seater
(1985) calculates marginal tax rates using data on actual tax payments and income
earned by individuals. The standard formula for his marginal tax rate is:
MTR j =
T j T j 1
Y j Y j 1
(4.1)
where j is an index for income brackets; MTRj is the marginal tax rate for bracket j; Tj is
the tax paid by an individual in bracket j; and Yj is the income earned by an individual in
bracket j.
The average marginal tax is then computed as the weighted average of MTRj
using fractions of total income in each bracket as weights. Seater also calculates
marginal social security tax rates and average marginal social security tax rates from
these. The overall marginal tax rate is then calculated as the sum of the different
average marginal tax rates.
Koester and Kormendi (1989) use a simpler, more direct, method of calculating
marginal tax rates: the overall total tax revenue (TAXREV) is regressed against GDP
over time (t) as follows:
TAXREVt = a 0 + a1GDPt + et
(4.2)
The coefficient on GDP in the above linear equation approximates the increase
in tax revenues associated with an increase in GDP, and is considered as the average
marginal tax rate in the economy.
128
Seaters (1985) and Koester and Kormendis (1989) marginal tax rate measures
differ substantially in that Seater looks at statutory tax rates while Koester and
Kormendi look at actual taxes paid. Nonetheless, either of these measures still appears
superior to Gwartney et al.s measure since marginal tax rates for individuals in the
economy are averaged instead of taking top marginal rates only.
However, these
marginal tax rates also ignore the fact that different marginal tax rates apply to different
incomes from different sources, i.e. a different marginal tax rate applies to income from
labour, and another rate for income from capital.
Joines (1981) recognises this flaw and computes different marginal tax rates that
apply to income from labour (e.g. compensation of employees) and income from capital
(e.g. rental income of persons; corporate profits), adjusted for miscellaneous income
(business and professional; farm). Similarly, Mendoza et al. (1997) (extending the
earlier work of Mendoza et al. 1994) compute marginal tax rates on factor incomes
(labour and capital) and consumption, taking into consideration the net effect of
different countries policies on credits, exemptions and deductions; the effect of taxes
that are not filed with individual tax returns (e.g. social security contributions, etc.); and
information on statutory tax rates and income distribution in each tax bracket.
The marginal tax rate measures that Joines and Mendoza et al. propose seem
superior to other measures, but the methods for calculating these marginal tax rates
require data that are available for a limited group of countries only, usually the OECD
countries.73
For instance, for the Mendoza et al. measure, data on the different
categories of taxes are required, and also data on each countrys distribution of income
consistent with the income tax schedule in that country.
Based on the preceding
discussion, it is acknowledged that the analyses in this chapter could be improved with
the inclusion of a marginal tax rate variable. However, for lack of reliable time-series
data on marginal tax rates, the average tax rates are used instead in this chapter.
73
Joines (1981) computes marginal tax rates for the United States only, while Mendoza et al. (1997)
compute marginal tax rates for 18 OECD countries only.
129
In terms of openness, we use Tradetax, which is the ratio of taxes on
international trade to the total trade sector (i.e. exports plus imports), to proxy for the
degree of openness of an economy.
Essentially, Tradetax is the average tax on
international trade. Ideally, we would have preferred also to examine how our results
would be affected if we use other measures of openness (e.g. the black market exchange
rate premium (Black) as we did in Chapter 3; however, annual time-series data for Black
are not available.74,75 Rodriguez and Rodrik (2000) also show that of the many different
indicators of trade restrictions used in previous empirical studies (e.g. Dollar, 1992;
Sachs and Warner, 1995; Ben-David, 1993, and Edwards, 1998), the simple average of
taxes on imports and exports, such as our Tradetax measure, does a good job of
measuring the degree of openness of an economy, even though results of their study
find no conclusive evidence that trade restrictiveness has a detrimental effect of
growth.76
Table 4.1 provides a detailed description of the data used in the following
analyses.
74
Black market exchange rate data are available from Barro and Lee (1994), but these are available at
five-yearly intervals only. The World Currency Yearbook (Cowitt, various years) also contain
information of the black market premia. However these are not available consistently for the entire time
period considered in this study. In addition, based on Rodriguez and Rodrik (2000), the black market
premium is not considered a good measure of openness because it is also a proxy for many other
variables, e.g. corruption, unreliable bureaucracy, etc., that are not necessarily associated with openness.
75
We do not consider the use of the Size variable from Gwartney et al. (1996) in this chapter due to the
concerns about the reliability of these data as discussed in Chapter 3, Section 3.3.1.
76
Rodriguez and Rodrik (2000) argue that the diverse results of previous empirical studies are due to
either mis-specification of the openness proxy, or from the use of proxies for openness that proxy for a
wide range of policy and institutional differences that are not necessarily related specifically to trade
policy or openness.
130
Table 4.1
Data for multivariate cointegration and causality analysis
Variable
Definition
Period covered and total

no. of observations
available
Expected effect on
TFPa
Tradetax
Taxes on international trade as

a percentage of total trade
sector (exports plus imports)
Indonesia: 1972-95 : 24
Korea: 1972-95 : 24
Malaysia: 1972-95 : 24
Singapore: 1972-95 : 24
Thailand: 1973-95 : 24
Negative
Indonesia: 1972-95 : 24
Korea: 1972-95 : 24
Malaysia: 1972-95 : 24
Singapore: 1972-95 : 24
Thailand: 1973-95 : 24
Negative, positive or
zero
Indonesia: 1973-95 : 23
Korea: 1972-95 : 24
Malaysia: 1972-95 : 24
Singapore: 1972-95 : 24
Thailand: 1973-95 : 24
Negative, positive or
zero
Indonesia: 1973-95 : 23
Korea: 1972-95 : 24
Malaysia: 1972-95 : 24
Singapore: 1972-95 : 24
Thailand: 1973-95 : 24
Positive
Source: IMF-GFS yearbooks,

various issues
Domtax
Taxes in the domestic

economy (including taxes on
income, profit, payroll and
manpower, property, social
security contributions, and
taxes on domestic goods and
services) as a percentage of
total GDP
various issues
Noncomp
Total government expenditure

on: general public services,
defence, and social security
and welfare, as a percentage of
total GDP
various issues, ADB Key
Indicators for Developing
Member Countries, various
issues for Malaysia
Comp
Total government expenditure

on: health, education and
economic services, including
expenditures on research,
roads, other transport and
communication services, as a
percentage of total GDP
various issues, ADB Key
Indicators for Developing
Member Countries, various
issues for Malaysia
continued over page
131
Variable
Definition
Period covered and total

no. of observations
available
Expected effect on
TFPa
Human
Labour quality index that

measures the quality of the
workforce based on a 7%
return to each additional year
of schooling, based on an
improved version of the Barro
and Lee (1994) data on
education.
Indonesia: 1960-96 : 37
Korea: 1960-96 : 37
Malaysia: 1960-96 : 37
Singapore: 1960-96 : 37
Thailand: 1960-96 : 37
Positive
Source: Collins and Bosworth

(1996)
TFP
Total Factor Productivity

levels
Indonesia: 1960-96 : 37
Korea: 1960-96 : 37
Malaysia: 1960-96 : 37
Singapore: 1960-96 : 37
Thailand: 1960-96 : 37
Source: calculated from new

data from Collins and
Bosworth (1996)
Note:
a
Please refer to section 4.2 for the reasons for the hypothesised signs.
4.3
TFP levels for the HPAEs

TFP levels are calculated from a levels accounting exercise similar to that of
Hall and Jones (1999) but without human capital in the underlying production function.
As pointed out in Chapter 3, we exclude human capital in our underlying production
function because its inclusion in the underlying production function affects the final
estimates of TFP and we propose that, from a theoretical perspective, any effect of
human capital on overall economic growth is actually transmitted via its effects on TFP
(see Section 3.2 for details). Our levels accounting equation is expressed as:
K
y t = t
Yt
(1 )
At
(4.3)
132
where,
yt
Kt
Yt
At
= Y L , level of output per worker at time t ,
= capital - output ratio at time t ,
= an index for total factor productivity (TFP ) at time t and,
is a parameter that measures the elasticity of output with respect to capital.

A number of empirical studies show that the estimated level of TFP is highly
sensitive to (e.g., Dowling and Summers (1998), Senhadji (1999)). With this in mind,
different values of were used to calculate TFP levels. Senhadji reports that estimates
of range from 0.13 to 1.00 in levels, but on average are 0.55.77 Based on this, we use
estimates of equal to 0.2, 0.33, 0.4, 0.5 and 0.6, in calculating different TFP levels.78
Using the latest available Collins and Bosworth (1996) dataset (hereafter CBnew),79 we calculate TFP levels for the five HPAEs (see Appendix H). Since the data
used to calculate the TFP levels are measured in the local currencies of each country,
comparing TFP levels across countries would be meaningless.
To enable TFP
comparison across countries, compounded TFP growth rates for two time periods are
also calculated. The first period for which growth rates are calculated is from 1960 to
1996, the longest time period of available data. The second period is from 1972 to
1996, the time period for which we do the succeeding analyses because the data for
other relevant variables are available only for this span of time. Table 4.2 presents the
TFP growth rates calculated from the TFP levels of the five HPAEs.
From equation (4.3), an estimate of equal to 1.00 would indicate that the growth in output due to the
growth in the capital-output ratio is undefined, or that a minute change in the capital-output ratio would
lead to a very small growth in output, which is implausible based on economic theory.
77
As we shall see in the results of the TFP calculations, the use of an estimate of greater than or equal
to 0.6 to calculate TFP levels results in negative TFP growth rates. This is a result that we find highly
implausible based on economic intuition, so we avoid using estimates of greater than 0.6 even though
Senhadji (1999) has found that some countries may have values of greater than this.
78
79
The CB-new dataset provides actual labour force data (in the previous CB dataset, labour force figures
have been indexed such that the labour force in year 1 (1960) is set equal to one in all countries), and a
new set of human capital indices has been calculated from a later, improved version of the Barro and Lee
data on education (Barro and Lee, 1996). The time series data have also been extended to 1996.
133
Table 4.2
TFP growth rates for selected HPAEs

= 0.2
Average : 1960-1996
Indonesia
2.81 (5)
Malaysia
3.16 (4)
Singapore
3.99 (3)
South
4.66 (1)
Korea
Thailand
4.16 (2)
Average : 1972-1996
Indonesia
3.19 (5)
Malaysia
3.35 (4)
Singapore
3.98 (3)
South
5.02 (1)
Korea
Thailand
4.56 (2)
= 0.33
2.21
2.46
2.95
3.73
(5)
(4)
(3)
(1)
TFP growth rate

= 0.4
1.80
1.97
2.24
3.10
(5)
(4)
(3)
(1)
= 0.5
1.01
1.05
0.89
1.89
(4)
(3)
(5)
(2)
= 0.6
-0.17
-0.34
-1.12
0.08
(3)
(4)
(5)
(2)
3.41 (2)
2.91 (2)
1.91 (1)
0.49 (1)
2.29
2.68
3.36
4.12
1.68
2.22
2.93
3.51
0.50
1.34
2.12
2.34
-1.25
0.02
0.90
0.60
(5)
(4)
(3)
(1)
4.14 (2)
(5)
(4)
(3)
(2)
3.86 (1)
(5)
(4)
(3)
(2)
3.30 (1)
(5)
(4)
(2)
(3)
2.48 (1)
Note:
Numbers in parentheses show HPAE ranking based on TFP growth rates, with 1 = fastest and 5
= slowest.
Consistent with the predictions of levels and growth accounting, we see in Table
4.2 that the larger the assumed share of capital in GDP (), the lower the TFP level (see
Appendix H) and TFP growth rate.
This is because increasing increases the
contribution of capital to output growth, but decreases the contribution of TFP

(Senhadji, 1999).
Table 4.2 also shows that depending on the estimated level of used and which
time period is considered, different countries appear to have the fastest or slowest TFP
growth rate. For instance if the estimated is either 0.2 or 0.33, then TFP grows fastest
in South Korea, followed by Thailand, Singapore, Malaysia and Indonesia regardless of
which time period is considered. However, for s greater than 0.33, the time period
considered appears to be a crucial factor in determining which country has the fastest or
slowest TFP growth rate. If the estimated level of is 0.4, South Korea has the fastest
TFP growth rate in the 1960 to 1996 period, but Thailand for the 1972 to 1996 period.
134
The country with the slowest TFP growth rate in either time period is Indonesia.
If the estimated level of is 0.5, again, Thailand appears to have the fastest, and
Indonesia the slowest TFP growth rate for the 1972 to 1996 period, but, for the 1960 to
1996 period, Thailand appears to have the fastest TFP growth rate, and Singapore the
slowest. Finally for an estimated level of equal to 0.6, Thailand has the fastest TFP
growth rate in both time periods. Note that Singapore has the slowest TFP growth rate
in the 1960 to 1996 period, but is second fastest in the 1972 to 1996 period. Note also,
that the ranking we make is conditional on the HPAEs having the same level of .
Senhadji (1999) however, suggests that can vary considerably across the HPAEs.
Nonetheless, consistent with the practice in many growth accounting studies, we assume
that the HPAEs as a group have the same level of .
We do not put much faith in the results obtained when we use an estimated
equal to 0.6 in calculating the TFP levels and growth rates. Recall that TFP is supposed
to measure the level of technology and efficiency in an economy. A negative TFP
growth rate, therefore, especially over a long period of time, would imply that there may
not have been any technological improvement at all over the entire period studied, or,
that while there may have been some technological improvement, the level of efficiency
has regressed.
These implications seem implausible considering the fact that the
HPAEs have grown and developed so much in the past thirty years.
Even if a
significant proportion of their growth can be attributed to factor accumulation (Kim and
Lau, 1994, 1995; Young, 1992, 1994, 1995; Collins and Bosworth, 1996; Dowling and
Summers, 1998 etc.), to say that technology and efficiency in these countries have not
advanced or have even regressed (as evidenced by negative TFP growth rates) is rather
questionable.
How reliable are the TFP levels and growths rates that we have calculated?
Very few empirical studies have so far calculated TFP levels that are comparable to
ours. For instance, Hall and Jones (1999) compute TFP levels for 127 countries for the
period 1988, but their TFP levels cannot be compared to ours, even for the 1988 figure,
because the data they used were in constant, 1985 international dollars, while we have
used data in constant, 1987 local currencies. Hall and Jones have also included a human
135
capital variable in their production function, while we have not. The TFP levels that
Nehru and Dhareshwar (1994) estimate, based on an error correction model without
human capital, are the closest TFP levels to which we could compare our results
because they also used data in 1987 local currencies, but they have not provided
detailed (yearly) estimates of their TFP levels.
Senhadji (1999) also estimates TFP
levels, but detailed results are also not available. For this reason, we can only compare
the compounded growth rates of TFP we have calculated from Table 4.2 to the
estimated TFP growth rates in selected empirical studies (see Table 4.3). We report
only our TFP growth rates calculated from an assumed that is similar to the assumed
in the other studies.

As seen in Table 4.3, the TFP growth rates according to various empirical
studies are very different. We calculate TFP growth rates from a levels accounting
exercise; the World Bank (1993a), Nehru and Dhareshwar (1994) and Young (1994)
derive TFP growth rates from regressions; and Collins et al. (1995), Dowling and
Summers (1998), and Kawai (1994) obtain their TFP growth rates from growth
accounting. We exclude human capital from our underlying production function, while
the World Bank (1993a) and Collins et al. (1995) include human capital in their
respective production functions. As pointed out earlier, we question the results of
empirical studies that report negative TFP growth rates (World Bank, 1993a, Nehru and
Dhareshwar, 1994 and Collins et al., 1995). We are also uncomfortable with the results
presented by Dowling and Summers (1998) because there appears to be no clear
relationship between the TFP growth rates they calculated and the estimate of used.
As mentioned earlier, the theory behind levels and growth accounting suggests
that increasing the assumed share of capital in output () increases the contribution of
capital to output, and consequently decreases the contribution of TFP, so the TFP
growth rate should also be lower. This pattern of a large and a low TFP growth rate
however is not apparent in the results of Dowling and Summers (1998). Young (1994)
finds that TFP growth is fastest in South Korea and Thailand in the 1970s to 1990s
period, a result similar to ours. However, we find that Singapores TFP growth rate is
136
Table 4.3
1960s
1990s
Indonesia
Malaysia
Singapore
South Korea
Thailand
1960s
1990s
Indonesia
Malaysia
Singapore
South Korea
Thailand
1970s
1990s
Indonesia
Malaysia
Singapore
South Korea
Thailand
Estimates of TFP growth rates from selected empirical studies

Table 4.2
Table 4.2
World Bank World Bank Nehru and

(1993)a
(1993) b
Dhareshwar
(1994)c
1960-96
1960-96
1960-1990
1960-1990
1960-1990
Collins,
Bosworth
and Chen
(1995)d
1960-1992
= 0.33
= 0.4
Regression
Regression
Regression
= 0.4
2.21
2.45
2.95
3.73
3.41
1.80
1.97
2.24
3.10
2.91
1.3
1.1
1.2
3.1
2.5
-0.8
-1.3
-3.0
0.2
0.6
0.19
-0.18
-0.61
0.71
0.09
0.40
-0.40
0.60
1.40
1.30
Dowling and Dowling and Dowling and Dowling and Dowling and Dowling and
Summers
Summers
Summers
Summers
Summers
Summers
(1998)e
(1998)e
(1998)e
(1998)f
(1998)f
(1998)f
1961-1995 1961-1995 1961-1995 1961-1995 1961-1995 1961-1995
= 0.3
= 0.35
= 0.4
= 0.3
= 0.35
= 0.4
1.4
1.6
2.9
1.1
1.7
1.0
2.9
4.1
2.6
2.7
1.2
1.7
3.2
1.1
1.5
2.3
1.7
2.9
2.0
1.0
2.6
1.8
3.1
2.1
0.8
3.0
1.8
3.2
2.1
1.0
Table 4.2
Table 4.2
Table 1
1972-1996
1972-1996
1972-1996
Young
(1994)
1970-1985
Kawai
(1994) g
1970-1990
= 0.33
= 0.4
= 0.5
= 0.45
=?
2.29
2.68
3.35
4.12
4.14
1.68
2.22
2.93
3.51
3.86
0.50
1.34
2.12
2.34
3.30
1.0
0.1
1.4
1.9
1.49
1.60
1.15
1.75
1.90
Notes:
Estimated TFP growth rates from a production function regression of 87 economies.
b
Estimated TFP growth rates from a production function regression of high income economies
only.
c
Based on an error-correction model without human capital.
d
Adapted from Drysdale and Huang (1997).
e
TFP growth estimates using Nehru and Dhareshwar capital stock and World Bank output.
f
TFP growth estimates using King and Levine (1994) capital stock and Summers and Heston
output.
g
Kawai (1994) reports TFP growth rates for the sub-periods 1970-80 and 1980-90, so the
estimates above are re-calculated as an average of these two sub-periods for comparison to the
estimates in the other studies.
a
137
not the lowest among the HPAEs, but Young finds Singapores growth rate to be the
slowest and hardly contributes to total output growth. Youngs TFP growth rates for
Singapore differ from ours because he obtains these by regressing growth of output per
worker on a constant term and growth in capital per worker, while we use a levels
accounting exercise to arrive at our TFP growth rates. Young notes that Singapores
TFP growth rate would be higher if computed from a growth accounting exercise, with
an assumed share of capital in output of 0.35.
Similar to Youngs results, Kawai (1994) finds that South Koreas and
Thailands TFP growth rates are the fastest among the HPAEs considered, and
Singapores the slowest. We however cannot compare our results with Kawais with
ease because Kawai has not been explicit about the estimate of he uses in his growth
factor analysis.
In general, there seems to be little similarity between the TFP growth rates of
the HPAEs we calculate in this study with the TFP growth rates in the empirical studies
cited above primarily because of the manner in which TFP growth rates are calculated.
Despite this, we maintain that the TFP growth rates we calculate from the TFP levels
accounting exercise are feasible because the results are consistent with the predictions
of levels and growth accounting, and in particular, with reference to the relationship
between s and TFP. In the succeeding analyses, we use the TFP levels that we
calculate under the assumption that the share of capital in output () is 0.4, which leads
to sensible estimates of TFP levels. = 0.4 has been commonly used in other growth
accounting studies of the HPAEs and other developing countries (e.g. Kim and Lau,
1995; Collins et al., 1995; Collins and Bosworth, 1996; Dowling and Summers, 1998).
In the next section, we describe the general method to be adopted for causality
testing including the appropriate treatment of non-stationary and cointegrated variables.
138
4.4
The relationship between the Granger causality concept and cointegration
4.4.1
The Granger causality concept
The degree of linear association of one variable on other variables can be

determined through simple regression analysis; however, this does not necessarily imply
causation. If we then want to determine whether government intervention, the degree of
openness or human capital causes TFP; or whether TFP causes these other variables,
we need to consider these relationships within the concept of Granger causality.
Consider for example, two time series: X1 and X2. Causality from X1 to X2 in the
concept of Granger (1969) means that using lagged values of X1, together with lagged
values of X2 will lead to better forecasts of X2, compared to using lagged values of X2
alone. That is,
i =1
i =1
X 2t = i X 2t i + i X 1t i + u t
(4.4)
where k is the lag length, often, but not always, chosen arbitrarily.80 If i = 0 (for all i =
1, 2, , k), then, X1 does not Granger-cause X2. Note, however, that some authors (e.g.
Leamer, 1985), prefer the use of the term precedence or informativeness rather than
causality as Grangers test does not really test for causality as it is usually understood;
instead, it only tests whether one variable precedes another.
The concept of Granger causality essentially hinges on the use of time-series

data. Related to this is the importance of investigating the stationarity properties of the
data being studied. This is because, if the series are non-stationary (as many economic
variables have been found to be), then coefficient estimates based on ordinary least
squares (OLS) regressions may be biased and inconsistent; i.e. the regression results can
be spurious. Thus, if the variables are non-stationary and cointegrated, then this needs
80
The bivariate equation (4.4) can easily be adjusted to accommodate other variables and reparameterised as an error correction representation as will be evident later in Section 4.7.
139
to be properly taken into account and estimated accordingly; otherwise, incorrect
conclusions about causality or non-causality may be made (Granger, 1988).
4.4.2
Cointegration, error correction model (ECM) and causation
The relationship between Granger causality (which is concerned with short-run

forecastability) and cointegration (which is concerned with long-run equilibrium) is best
considered within an error correction model (ECM), but first, we define cointegration.
The concept of cointegration made popular by Engle and Granger (1987) has attracted
considerable attention in the empirical research on economic growth. In this chapter we
apply the concept of cointegration to estimate directly and test for the existence of longrun relationships among TFP, openness, government and human capital. Cointegration
describes a situation where individual time series that are integrated (i.e., are nonstationary) form linear combinations that are stationary. The presence of cointegrating
relations can be examined within the framework of error correction models. ECMs,
sometimes also referred to as equilibrium correction models (Hendry, 1995), are closely
related to the concept of cointegration. The idea behind ECMs is that the short-run
dynamics (which are reflected in the coefficients of the differenced variables) or the
departures from the long-run relationships between cointegrated variables can be
described by an error correction term. A long-run relationship among the variables (i.e.,
the variables are cointegrated) is determined if the coefficient of the error correction
term in the ECM equation is negative and statistically significant.
To illustrate the concept, let Zt follow a pth-order vector autoregressive (VAR)

process for the N x 1 vector Zt,
Z t = i Z t i + e t
(4.5)
i =1
where et is white noise, i.e., iid(0,2). To express equation (4.5) as an ECM, we rewrite it as:
140
p 1
Z t = Z t 1 + i Z t i + et
(4.6)
i =1
The coefficients and i for i = 1,2, , p1 in the ECM representation

(equation (4.6)) are determined from coefficients i for i = 1, 2, , p of the VAR
representation (equation (4.5)).
In this ECM (equation (4.6)), Zt must be Granger caused by Zt-1. This implies
that, for a pair of variables to reach a long-run equilibrium, some causation between
them that provides the necessary dynamics must be present (Granger, 1988). In other
words, where there is cointegration, Granger-causality (at least in one direction) must be
present.
We therefore test first whether or not our variables of interest are stationary and
whether or not there is a long-run equilibrium relationship among the non-stationary
variables (i.e., non-stationary variables are cointegrated). These pre-tests are essential
in determining whether cointegration exists before actually establishing the direction of
causality in order to reduce the possibility of identifying spurious causality (Oxley,
1993; Toda and Phillips, 1993, 1994).
Essentially, we are following a three-step
procedure to test for causality. First we examine the nature of all the data we use
through unit root tests. Second, we conduct cointegration tests on the non-stationary
variables, and, third we proceed with Granger causality testing. This is similar to Engle
and Grangers (1987) two-step procedure, however, we add another step, the unit root
tests, and also adopt different methods for the unit root and cointegration tests, which
are discussed in the following sections.
Causality tests may also be carried out within a vector autoregressive (VAR)
framework rather than within an ECM framework. For instance, Toda and Yamamoto
(1995); propose a procedure that uses a VAR representation in levels, choosing the lag
length k, based on usual procedures, then estimating a (k + dmax)th order VAR where
dmax is the maximal order of integration suspected in the process (LA-VAR). Kurozomi
and Yamamoto (2000) note that the LA-VAR method suffers from inefficiency because
141
of the artificially augmented lag, and modify Toda and Yamamotos (1995) procedure
and propose an alternative procedure, modified lag augmented VAR (MLA-VAR). The
MLA-VAR procedure essentially involves dividing the sample into two sub-periods and
correcting for the bias associated with the OLS estimator using a grouped jackknife
method. Kurozomi and Yamamoto find that the MLA-VAR is superior to the LA-VAR
in terms of improved empirical size without losing power.
The Monte Carlo
experiments Ng (2000) carries out to investigate the power and size performance of the
ECM, LA-VAR and MLA-VAR procedures indicate that the ECM procedure is superior
to the other two procedures in terms of power, but suffers more from large size
distortions. In this study, we opt to test for causality within an ECM framework, as it is
found to be more powerful, while acknowledging that we are trading-off size stability
for power.
4.5
Analysis of the time-series properties
4.5.1
Unit root testing procedure
Visual inspection of our variables: TFP, Tradetax, Domtax, Noncomp, Comp

and Human, suggests that most of the series are likely to be non-stationary. In order to
formalise this perception, unit root tests are undertaken to determine the order of
integration (i.e., the number of times the series needs to be differenced in order to be
stationary) of the time series. There are several procedures for testing for unit roots,
including those developed by Fuller (1976), Dickey and Fuller (1979, 1981), Said and
Dickey (1984), and Phillips (1987). The most frequently used unit root test in the
empirical literature is the augmented Dickey-Fuller (ADF) test.
We however, adopt the unit root testing procedure suggested by Holden and
Perman (1994) for the logs of all the variables in levels and in first differences, with the
I(2) versus the I(1) test carried out first followed by the I(1) versus I(0) test (Dickey and
Pantula, 1987; Pantula, 1989). The first step is to use the statistic 3 to test the null H0:
(, , ) = (, 0, 0) against the alternative H1: (, , ) (, 0, 0) based on the
autoregression of differences, with constant and trend:
142
y t = + t + y t 1 + i y t i + u u
(4.7)
i =1
This drawn-out sequence of tests is preferable to the usual ADF testing

procedure as it attempts to trade-off the loss of power (from having too many irrelevant
coefficients in the model) for ensuring similarity with the ADF tests and having
plausible alternatives (under H1).
If the null hypothesis H0: (, , ) = (, 0, 0) is rejected, there are three
possibilities: (a) the series has a unit root ( = 0) with trend ( 0); (b) the series has no
unit root ( 0) and no trend ( = 0); or (c) the series has no unit root ( 0) but is
with trend ( 0). The next step then is to test ( = 0) using the t-statistic from the
estimated equation (4.7), with critical values from standard normal tables. If ( = 0) is
not rejected, then we can conclude that the series has a unit root and a trend.81 If ( = 0)
is rejected, this leaves possibilities (b) and (c), where ( 0) in both cases, i.e., the
series has no unit root, but could have a trend. A standard t-test for ( = 0) can be done
to determine whether the series has a trend or not. If ( = 0) is rejected, then the series
has no unit root but with a trend; but if ( = 0) is not rejected, then the series has no unit
root and no trend. In either case, we can also continue to perform standard t-tests to
determine whether or not the series has an intercept, i.e. test for ( = 0).
If the null hypothesis H0: (, , ) = (, 0, 0) is not rejected, then the series has
a unit root ( = 0) with no trend ( = 0), but with a possible drift. A further test on ( =
0), assuming ( = 0) using a t-statistic with non-standard critical values82 is made to
establish that the series does have a unit root. Once this is established, a 2 test for (,
, ) = (0, 0, 0) is carried out. If this is rejected, then we conclude that the series is a
random walk with drift; but if it cannot be rejected, then the series is a random walk
without drift. In testing whether the series is I(2) vs. I(1), (i.e., for the unit root tests on
81
Campbell and Perron (1991) point out that this is not a sensible result based on a priori grounds,
because it implies that the series is actually exploding because of the presence of a trend.
143
first differences), additional tests based on equation (2) with ( = = 0) were also
carried out because it is possible that the mean of the differenced series may be zero.
4.5.2
Choosing the appropriate lag length
Test results often depend critically on the number of lagged differences included
in the regression equation. The reason for the augmentation of the lagged differences is
to ensure that the errors are uncorrelated. The consequences of over-specified lag
lengths would be similar to the problems associated with the inclusion of irrelevant
variables (in general linear models), i.e., the variances tend to be larger than necessary,
while under-specified lag lengths (similar to the case of omitted variables in general
linear models) result in biased and inconsistent estimates, but with a smaller variance
(Gujarati, 1995). Thus, if arbitrary lag values are specified, the subsequent statistical
inferences are likely to be invalidated. In response to this problem, several procedures
have been proposed for allowing the data to determine the appropriate length of the
distributed lag, e.g., based on Akaikes (1973) (AIC) or Schwarzs (1978) (SBC)
information criteria (White, 1993); or on Campbell and Perrons (1991) and Halls
(1994) general to specific (GS) testing strategy, among others.
As mentioned previously, choosing a lag length arbitrarily, or independently of

the data, is likely to have serious effects on the test results, thus pointing to the
importance of selecting the lag length appropriately with the use of data-based model
selection procedures (e.g. AIC, SC, GS, among others). Based on the general tendency
of the GS to choose models of higher order than the true model, the GS procedure as
proposed by Campbell and Perron (1991) and Hall (1994) is used in this study,
primarily because, as mentioned previously, the consequences of including a larger
number of lags than is necessary are less serious than when lag lengths are underspecified (Gujarati, 1995).
82
Non-standard critical values are available in Dickey and Fuller (1981), but the SHAZAM output
(White, 1993) reports McKinnon (1991) asymptotic critical values.
144
The procedure for the GS testing strategy suggested by Campbell and Perron
(1991) and Hall (1994) starts by estimating an autoregression of a large order, i.e., of
order pmax, where pmax is an upper limit on p, chosen a priori. If a significant statistic is
not obtained on the pth lagged difference term, then the order of the estimated
autoregression is reduced by one, iteratively, until a significant statistic is obtained. If
none is significant, then p = 0 is selected. In this study, considering the short span of
time series data available, pmax is set at 5.
4.5.3
Detailed unit root results in the levels and first differenced series for the logs of
TFP, Tradetax, Domtax, Noncomp, Comp and Human are presented in Appendix I. The
unit root results are based on 24 annual observations from 1972 to 1995 for each data
series for each country, except for the unit root results for the Noncomp and Comp
variables for Indonesia, which are based on 23 annual observations from 1973 to 1995
as data for these variables are available only for this period.
The unit root tests on the logs of TFP, Tradetax, Domtax, Noncomp and Comp
in first differences show that the null of a unit root is generally rejected for all the
HPAEs. This suggests that the series are stationary in first differences, but could still be
non-stationary in levels. By contrast the unit root tests on the log of Human in first
differences show that the null of a unit root is rejected for Indonesia and Singapore, but
is not rejected for Malaysia, South Korea and Thailand. This implies that the log of
Human is stationary in first differences but could be non-stationary in levels for
Indonesia and Singapore. The I(1) versus I(0) tests for the log of Human for Indonesia
and Singapore indicate that Human is stationary for Indonesia and non-stationary for
Singapore.
For Malaysia, South Korea and Thailand, the first differenced series of the log of
Human is non-stationary suggesting than Human is likely be integrated of order two (or
higher) for these HPAEs.
If correct, this result would have implications for the
subsequent cointegration analyses because of the presence of a potential I(2) variable on
145
the right-hand side of our regression equation. Why does Human appear to be an I(2)
series in these countries?
We offer two plausible explanations for this. The first explanation is related to
the way the Human series has been constructed. Lynde and Richmond (1993) show that
some economic series, specifically public capital in their study, may appear I(2), even
though the series is actually I(1), because of the way the series has been constructed. To
illustrate, Lynde and Richmond construct their public capital, G, as:
Gt = (1 )Gt 1 + Z t
0 < (1- ) < 1
(4.8)
where Zt is gross public investment. From equation (4.8), it follows that,

2 G t = G t 1 + Z t
(4.9)
and if Z is I(1) and the depreciation rate, , is positive but close to zero (Lynde and
Richmond use = 0.008), then, the ADF unit root test on G will likely fail to reject the
null of a unit root, and so G will appear I(2), even though from equation (4.8),
Gt = (1 )Gt 1 + Z t
0 < (1 ) < 1
(4.10)
so Gt is stationary and hence, G is actually I(1).
In our case with the Human capital index series, the analysis is less
straightforward. If we are to apply Lynde and Richmonds analysis to our case with the
Human series, we need to look closely at how Collins and Bosworth (1996) (our data
source) construct the series from Barro and Lees (1994) data on educational
attainment.83
83
The Collins and Bosworth (1996) human capital data that we use in this study are derived from a later,
improved vintage of the Barro-Lee data (i.e. from Barro and Lee, 1996), but since Collins and Bosworths
method of constructing the educational attainment data, particularly with respect to the use of a perpetual
inventory method, remains similar to when the original Barro-Lee (1994) data are used, we opt to analyse
the construction of the original data set instead here.
146
Collins and Bosworth simply adjust Barro and Lees data on educational
attainment with weights that represent returns (in terms of wage rates) to different levels
of educational attainment.
For now, we look at how Barro and Lee (1994) construct their educational
attainment data. Barro and Lee use national census data for about 40% of the possible
cells for their panel data set that cover over 100 countries observed at five-year intervals
from 1960 to 1985. Adult illiteracy rates are used to fill in cells where data for the noschooling category are unavailable. The remaining cells are filled using a perpetual
inventory method. The use of this perpetual inventory method in the construction of
part of the educational attainment data could be contributing to the Human series
appearing to be I(2) in some of the countries studied.
Let us examine Barro and Lees perpetual inventory procedure for the noschooling category (the procedure is analogous for the other schooling categories,
hence these are not discussed separately):
H 0t = (1 t )H 0,t 5 + [L 25 t (1 PRI t 15 )]
(4.11)
where,
H0t
= the estimated number of persons aged 25 and over who have no

educational attainment at time t;
H0t-5
= the estimated number of persons aged 25 and over who have no

educational attainment one period earlier, or at time t-5;
L25t
= the population aged 25 to 29 at time t;
(1-PRIt-15)
= the fraction of the population aged 25 to 29 who have no elementary

education (i.e. not enrolled in primary school 15 years before); and,
= the proportion of people aged 25 and over in year t-15 who did not
survive to year t; this can be considered the depreciation rate even
though it is more of a death probability that varies over time.
147
If this death probability, , is positive and close to zero, then, similar to the
case of public capital in Lynde and Richmond (1993), the ADF unit root test is likely to
fail to reject the null of a unit root on H0t, and, H0 may appear nearly I(2), even if it is
actually I(1). If this educational attainment data from Barro and Lee, exhibits I(2)
characteristics, then it is likely that such characteristics may be also manifested in the
Human series that were derived by Collins and Bosworth, so Human also appears I(2),
but may not actually be so.
The second explanation we propose is related to the way Collins and Bosworth
linearly interpolate the data to derive annual observations from the five-yearly
observations of the educational attainment data of Barro and Lee.
Collins and Bosworth are not explicit as to how they derive annual estimates of
their human capital index, when Barro and Lees (1994) data are available only at fiveyear intervals over the period 1960 to 1985.84 However, looking at the plot of the yearly
change in the human capital index for each country (see for example those for Malaysia,
South Korea and Thailand in Appendix J), it appears that Collins and Bosworth simply
use linear interpolation to derive annual estimates of their Human index. This may
therefore make the human capital index less useful in the time-series (and panel data)
exercises of the kind done in this study.
Hence a general caveat regarding the
interpretation of results relating to human capital is made here.
Linear interpolation also yields very smooth series over time, which is
characteristic of an I(2) series, but we know that the series is smooth primarily because
it is interpolated. The plots of the yearly changes in the Human series (see the examples
in Appendix J), show some distinct breaks in the series, which approximately
correspond to the five-yearly intervals in the underlying Barro-Lee data.
Several
authors (e.g. Perron, 1989; Reichlin, 1989; and Hendry and Neale, 1991) emphasise
that, in the presence of breaks in the time series, standard unit root tests tend to find in
favour of an I(d) series, when in fact the series is I(d-1).
84
The Barro and Lee (1996) data extend to 1996.
148
Based on the two explanations above, although our unit root test results show
that the log of Human appears to be I(2) for Malaysia, South Korea and Thailand, it is
justifiable to assume that this variable is more likely I(1), but artificially exhibits I(2)
characteristics. Hence, it is treated as an I(1) series for Malaysia, South Korea and
Thailand.
The unit root tests on the level series of the log of TFP for Indonesia, Malaysia,
Singapore and Thailand fail to reject the null of a unit root. For South Korea however,
the null of a unit root in the level series is rejected. We conclude that the log of TFP is
non-stationary (I(1)) for Indonesia, Malaysia, Singapore and Thailand, but is trend
stationary (I(0)) for South Korea.
The unit root tests on the log of the Tradetax series in levels show that the null
of a unit root is not rejected for Indonesia, Malaysia and Thailand. This implies that the
log of Tradetax is I(1) in these countries.85 For Singapore and South Korea, the null of
a unit root in the level series is rejected, implying that the log of Tradetax is stationary
(I(0)) for these two HPAEs, with Tradetax being trend stationary for Singapore, and
level stationary for South Korea.
The unit root tests on the log of Domtax in levels fail to reject the null of a unit
root for Indonesia, Malaysia and South Korea, implying that this variable is I(1) in these
countries. The series of the log of Domtax appears to be level stationary for Singapore,
and trend stationary for Thailand.
The unit root tests on the log of Noncomp show that this series in levels is I(1)
for Indonesia, South Korea and Thailand, but is level stationary for Malaysia and
Singapore.
85
Strictly speaking, variables that are in ratio form, e.g. Tradetax, Domtax, Noncomp, and Comp, cannot
be non-stationary, because they are bounded between 0 and 1. Nevertheless, it is possible that within this
bound, the series can best be locally approximated as non-stationary.
149
Finally, for the unit root tests on the log of Comp, this series in levels is found to
be non-stationary (I(1)) for Indonesia, Malaysia, Singapore and Thailand, but is level
stationary (I(0)) for South Korea.
Table 4.4 summarises the unit root test results by country. As we see, the orders
of integration of the variables are different in each country; hence we cannot generalise
the order of integration of any specific variable. More importantly, we also see that for
a particular country the variables that are to be included in the cointegration analyses do
not have the same orders of integration and this has important implications for the
results of cointegration tests. Including I(0) series in the regression equations increases
the number of cointegrating relations because each I(0) series is stationary by itself, and
so forms a cointegrating relation by itself (Harris, 1995). It is quite important then to
remember this point when analysing and interpreting the results of cointegration tests.
As for the I(2) series, the treatment of these in the regression equations is less
straightforward and the implications for the results of cointegration tests are more
complicated (Harris, 1995). Although Johansen (1995) has developed an approach that
specifically handles the inclusion of I(2) variables, for reasons mentioned previously,
the log of Human which appear to be an I(2) series for Malaysia, South Korea and
Thailand is actually more likely to be an I(1) series and will therefore be treated as such.
For South Korea, because log TFP, the dependent variable in our regression
equation, is I(0), we have two options. The first option is to difference all the I(1) series
(Domtax, Noncomp, and Human) to obtain I(0) series and then proceed to estimate
using standard regression techniques (e.g. OLS) for stationary variables. The second
option is to test for cointegrating relationships among South Koreas I(1) series
(Domtax, Noncomp and Human); however, interpretation of any cointegrating
relationships found would not be very meaningful in light of the specific objectives of
our study. For this reason, we choose option one for South Korea, presenting the results
in section 4.6.3 below. The next step in our empirical investigation involves applying
Johansens (1988) multivariate cointegration procedure in re-examining TFP and its
determinants.
150
Table 4.4
Variables
(in logs)
Summary of Unit Root Test Results

Indonesia
Malaysia
Singapore
South Korea
Thailand
TFP
I(1)
I(1)
I(1)
I(0)
I(1)
Tradetax
I(1)
I(1)
I(0)
I(0)
I(1)
Domtax
I(1)
I(1)
I(0)
I(1)
I(0)
Noncomp
I(1)
I(0)
I(0)
I(1)
I(1)
Comp
I(1)
I(1)
I(1)
I(0)
I(1)
I(1)
I(2)/I(1)
I(2)/I(1)a
Human
I(0)
I(2)/I(1)
Note:
a
Formal results suggest that the variable is possibly I(2), but is classified as I(1) on the basis of
the reservations regarding how the data are constructed by Collins and Bosworth (1996) (see
text).
4.6
Cointegration and error correction model (ECM) representation for the HPAEs
4.6.1 Johansens multivariate cointegration and ECM representation
Recall the ECM represention (equation (4.6)) in section 4.4.3,

p 1
Z t = Z t 1 + i Z t i + et
(4.6)
i =1
Johansen (1988) concentrates on this matrix, which is the long-run or

equilibrium impact matrix for the VAR representation of Zt. This matrix will be of less
than full rank if the vector is cointegrated. The rank of this matrix will equal the
number of cointegrating vectors. Johansen outlines three possible scenarios for the
matrix:
a)
If rank = N (full rank), any linear combination of Zt is stationary and

the autoregressive model can be estimated in levels.
b)
If rank = 0 (null matrix), the model should be estimated in differences.
c)
If rank = r, where 0 < r < N, implying that there are some linearly
independent stationary combinations of the Zts.
151
Imposing the appropriate restriction, due to the presence of cointegrating
relationships in Zt, equation (4.6) becomes:
p 1
Z t = Z t 1 + i Z t i + et
(4.12)
i =1
where the term Zt-1 defines the disequilibrium errors at time t-1 and are the
feedback coefficients. One advantage of Johansens procedure is that it tests for the
number of cointegrating relationships.
Johansen (1988) and Johansen and Juselius (1990) propose two likelihood ratio
tests for cointegration: a likelihood ratio (LR) trace test and a maximal eigenvalue (max) test. The LR trace test tests the null of r0 cointegrating relations (0 r0< N)
against the alternative of r0 +1 cointegrating vectors. The LR trace tests statistic is:
Trace r0 = T
ln(1 )
N
(4.13)
i = r0 +1
On the other hand, the -max tests the null of r0 cointegrating relations (0 r0<
N) against the alternative of r0 + 1 cointegrating relations, and is given by:
max,r = T ln r +1
0
r0 = 0 ,..., N 1
(4.14)
The distributions of the LR trace and -max test statistics are not given by the
usual chi-squared distributions, i.e., they have non-standard distributions, because Zt is
multivariate I(1). Johansen and Juselius (1990) report the relevant tables for the critical
values. Podivinsky (1990), as cited in Maddala and Kim (1998), finds that these critical
values based on the asymptotic distribution are inappropriate when the sample size is
100 or smaller. To address this problem, Reimers (1992), suggests adjusting the test
statistics to obtain satisfactory size properties, which involves multiplying the test
statistics by a factor (T-np)/T, where T is the sample size; n is the number of variables in
the VAR system; and p is the number of lags in the VAR system. Doornik and Hansen
152
(1994) however point out some concerns regarding the appropriateness of such
adjustment. In this study, we continue to use Johansen and Juselius (1990) critical
values, bearing in mind the aforementioned reservations. Also, if a low-order VAR
model is used, Toda (1995) shows that both the LR trace test and -max test tend to
spuriously find cointegration, but that the -max test is better than the LR trace test
when the power of the tests is low.
For both LR trace and -max tests, the sequence for the procedure starts by
testing the null hypothesis, r = 0. If the null is rejected, we continue to test r = 1, , r =
N-1 until the null is not rejected.
Johansens cointegration tests can depend critically on the number of lagged

terms included in the regression equation and which deterministic components
(intercept and trend) are included in the model.
However, in determining the
appropriate lag length, we require information on how the deterministic components

(trend and intercept) are treated in the model. The different treatments of deterministic
components can be dealt with by five models, only three of which are realistic
(Johansen 1992).86 Model 2 assumes restricted intercepts and no deterministic trends.
It is appropriate only when the mean of the differences of the variables are zero. Model
3 assumes unrestricted intercepts and no deterministic trends and Model 4 assumes
unrestricted intercepts and restricted trends. As mentioned previously, determining the
appropriate lag length to use depends on the treatment of the deterministic components
under Models 2 to 4, but the choice between Models 2 to 4 requires information on the
appropriate lag length to use, thus creating a vicious circle.87
86
Model 1 assumes no deterministic components in the data or in the cointegration relations and Model 5
is an extension of Model 4 that allows for linear trends in the first difference of the variables and
quadratic trends in the levels of the variables. These models are generally not considered because they
are unlikely to exist in practice. For Model 1, it is unrealistic because a constant is often needed in the
cointegrating relationship to account for units of measurement, and Model 5 is unrealistic because it
assumes an infinitely increasing or decreasing rate of growth if the variables are in logarithmic form
(Osterwald-Lenum, 1992).
87
See Humphries (1999) for a more detailed discussion of this problem.
153
With these factors in mind, we begin our analyses by first selecting the
appropriate lag length based on the Akaike (AIC), Schwarz Bayesian (SBC) criteria and
a likelihood ratio (LR) test. For the AIC and SBC, the appropriate lag lengths are
selected by maximising the following log-likelihood functions over the different choices
for the maximum lag (Pesaran and Pesaran, 1997):88
()
~
= l ( )
AIC:
~
AIC l = l n p
SBC:
SBC l
1
2
p log n
~
where is the maximum likelihood estimator of , the number of freely estimated
parameters based on a sample of size n, and p dimension.
Because of the short span of time-series data available, the maximum possible
order for the ECM is only two. The strategy for selecting the appropriate lag length is
apparent in the first table of Appendices K.1 to K.4.
We then apply a joint test for the determination of cointegration rank and the
appropriate model for the deterministic components, using the LR trace test and the max test. The testing procedure for the joint test begins by applying the LR trace and
the -max tests on the most restricted model (rank = 0; Model 2). If the null in Model 2
is rejected, we test rank = 0 in Model 3 next, then in Model 4 if still rejected. If the null
rank = 0 is rejected in all the models, we continue to test the null rank = 1 in Model 2,
88
The choice of appropriate lag lengths based on the AIC and SBC can also be made by minimising the
following functions over the different choices for the maximum lag:
AIC: (RSS + 2K2)/T
SBC: (RSS + K(logT) 2)/T
where RSS is the residual sum of squares, K is the number of regressors, and T is the number of
observations (White, 1993).
154
and follow the same sequence until the LR trace and the -max tests fail to reject the
null (Johansen, 1992).
4.6.2
Testing for cointegration in the HPAEs.
We define the vector of variables for Indonesia, Malaysia, Singapore and Thailand
with variables in logarithmic form as:
Z = {TFP, Tradetax, Domtax, Noncomp, Comp, Human}
and estimate an ECM of the following form:

p 1
Z t = Z t 1 + i Z t i + et
(4.15)
i =1
where t = 23 for Indonesia (annual observations from 1973 to 1995) and 24 for
Malaysia, Singapore and Thailand (annual observations from 1972-1995). Note that the
sample size for the succeeding cointegration analyses is less than is desirable for the
Johansen method.
The detailed results, i.e. the results of the tests associated with choosing the
appropriate lag length, and the LR trace and -max tests for cointegration following the
testing procedures outlined in the previous sections are presented in Appendix K. From
these we find that there are at least two cointegrating vectors for Indonesia, Malaysia
and Thailand, and one cointegrating vector for Singapore.
We present below the
estimates of the cointegrating vector(s) in each country, in normalised form (all

normalised on TFP).
Indonesia: Model 2, r = 1
Cointegrating vector 1:
TFP
= 3.519 + 0.244Tradetax + 0.330Domtax 0.459Noncomp +

0.528Comp
155
Malaysia: Model 3, r = 2
TFP
= 8.061Tradetax + 24.061Domtax 4.837Comp 61.280Human
TFP
= 3.661Tradetax 2.005Domtax + 4.810Comp 14.614Human
Singapore: Model 4, r = 1
TFP
= 0.282trend + 5.548Comp 116.578Human
Thailand: Model 3, r = 2
TFP
= 0.588Tradetax 0.966Noncomp 0.321Comp + 4.464Human
TFP
= 0.712Tradetax 0.153Noncomp + 0.412Comp 0.788Human
One major problem of the Johansen cointegration procedure is that of properly

identifying the cointegrating relations if r > 1, as is the case for Malaysia and Thailand.
That is, the approach identifies the number of cointegrating relations present, but does
not identify the actual cointegrating relationship(s) (Maddala and Kim, 1998). In the
results presented above, the cointegrating vectors have been normalised in such a way
that TFP appears as the dependent variable, but this does not necessarily mean that this
is the true cointegrating relationship, as any combination of the variables can form a
cointegrating relationship. In situations like this, we should be able to turn to economic
theory and economic models to help identify the true cointegrating relationships.
However, this is difficult, lacking a sound TFP theory and model. We also notice that
none of the equations above appears to be a reasonable TFP equation. This is because
the signs of some of the parameters, particularly, Tradetax for the cointegrating vector 1
in Indonesia, Malaysia and Thailand; Human for the cointegrating vectors for Malaysia,
Singapore and Thailand, do not have the correct sign based on our initial hypotheses.
156
We attempt to attach some economic meaning to the results of our cointegration tests by
testing for Granger causality (see Section 4.7). For now, suffice to say that we have
determined long-run cointegrating relationships among our variables of interest. We
acknowledge however, that our cointegration test results are quite fragile due to the
small-sample-size problem associated with the Johansen test noted earlier. Toda (1995)
also notes that both the LR trace and -max tests tend to spuriously find cointegrating
relations when the sample size is small and a low-order VAR model is used.
4.6.3
OLS estimation results for South Korea
For South Korea, where the I(1) variables have been differenced, so that all the
variables are treated as stationary, we estimate the following equation by OLS:
ln TFPt = 0t + 1t ln Tradetax + 2t ln Domtax + 3t ln Noncomp
+ 4t ln Comp + 5t ln Human + t
(4.16)
where the variables are as defined in Table 4.1, and t = 24 (annual observations from
1972 to 1995). The results are presented below (t-ratios are in parentheses):
ln TFPt = 17.508 + 0.136 ln Tradetax + 0.161 ln Domtax
(12.886) (0.892 )
(0.154)
(4.17)
0.326 ln Noncomp + 0.869 ln Comp + 8.728 ln Human

( 0.611)
(1.971)
(0.772)
R2 = 0.242
R2 = 0.020
The above results indicate that government complementary expenditures (Comp)

is a statistically significant positive determinant of South Koreas TFP at the 10% level
of significance. This result however cannot be considered robust since the whole
regression has an adjusted R-squared of only 0.020.
157
4.7
Testing for Granger causality
Given the results that we find at least one cointegrating relationship for
Indonesia, Malaysia, Singapore and Thailand, we conduct ECM-based causality tests
using the Johansen cointegrating vectors identified for each HPAE above.
For
Indonesia, the error correction representation of the model for the five I(1) variables is
as follows:
p 1
ln TFPt = 0 ln TFPt 1 + 1i ln Tradetax t i

i =1
p 1
p 1
i =1
i =1
+ 2i ln Domtaxt i + 31 ln Noncomp t i
(4.18)
p 1
+ 4i ln Comp t i + ln Humant + 61t 1 + u1t

i =1
p 1
ln Tradetax t = 0 ln Tradetax t 1 + 1i ln TFPt i

i =1
p 1
p 1
i =1
i =1
+ 2i ln Domtaxt i + 3i ln Noncomp t i
(4.19)
p 1
+ 4i ln Comp t i + ln Humant + 61t 1 + u 2t

i =1
p 1
ln Domtaxt = 0 ln Domtaxt 1 + 1i ln TFPt i

i =1
p 1
p 1
i =1
i =1
+ 2i ln Tradetax t i + 3i ln Noncomp t i
(4.20)
p 1

i =1
p 1
ln Noncomp t = 0 ln Noncomp t 1 + 1i ln TFPt i

i =1
p 1
p 1
i =1
i =1
+ 2 k ln Tradetax t i + 3i ln Domtax t i
p 1

i =1
(4.21)
158
p 1
ln Comp t = 0 ln Comp t 1 + 1i ln TFPt i

i =1
p 1
p 1
i =1
i =1
+ 2i ln Tradetaxt i + 3k ln Domtaxt i
(4.22)
p 1
+ 4i ln Noncomp t i + ln Humant + 61t 1 + u 5t

i =1
where p = 2, the lag length of variables TFP, Tradetax, Domtax, Noncomp, and Comp
(all in logarithmic form) and 1t-1 is the one period lagged value of the error correction
term. The coefficients of each of the regressors capture the short-run dynamics while
the coefficients of the error correction terms capture the adjustment to equilibrium.
Since the optimal lag length of the explanatory variables is one89 in each of the
regression equations, a standard asymptotic t-test on the coefficients of the explanatory
variables is used to test the null hypothesis of non-causality. A significant coefficient
rejects the null and we conclude that that particular variable Granger-causes TFP. For
bi-directional causality, the coefficients of the two variables being analysed in the
presence of other factors (i.e., other regressors) should be jointly significant.
For
example, bi-directional causality between TFP and Tradetax is determined if the 1 and
1 coefficients (in equations (4.18) and (4.19) respectively) are jointly significant. If
only the 1 coefficient is significant, then the causal relationship is from Tradetax to
TFP, and if only the 1 coefficient is significant, then the causal relationship is from
TFP to Tradetax. The error correction models (ECMs) for Malaysia, Singapore and
Thailand can be presented in a similar fashion to that of Indonesia.
Particular
adjustments are made however with respect to the specific cointegrating model and
number of cointegrating vectors identified and the stationarity of the variables for each
country series.
Detailed results of the ordinary least squares (OLS) estimation of the ECMs for
each of the variables for Indonesia, Malaysia, Singapore and Thailand are presented in
89
The optimal lag length test procedure adopted for the Johansen (1988) cointegration procedure in the
previous section identify an optimal lag length equal to two (i.e. p = 2), but since the causality model is
written in differenced form, the optimal lag length for lagged differences is effectively reduced to one.
159
Appendix L. Diagnostic tests for each of the estimations are carried out and results are
also presented in Appendix L.
The ECM estimation results for Indonesia indicate that none of the explanatory
variables Granger-causes TFP, although TFP appears to Granger-cause Comp in the
short-run. The error correction term in the TFP equation for Indonesia has the correct
(negative sign) but is insignificant, suggesting that the TFP equation does not actually
represent a cointegrating relationship. We find no evidence that government affects the
openness of the economy.
For Malaysia, Human and Noncomp are found to Granger-cause TFP in the
short-run, with Human positively influencing TFP, and Noncomp negatively influencing
TFP. Again, no evidence of bi-directional causality from TFP to these variables is
found. The second error correction term in the TFP equation for Malaysia is significant,
but is again positive, thereby casting doubt on the validity of the TFP equation as a
representation of a cointegrating relationship. We also find some interesting causal
relationships among the other variables under study; for instance, we find that Comp
and Human Granger-cause Tradetax and that the effect of Comp on Tradetax is
negative,90 but the effect of Human on Tradetax is positive.
For Singapore, we find that none of the explanatory variables included in the
TFP equation Granger-causes TFP, but find that TFP Granger-causes Comp, and
Human in the Comp equation and Human equation respectively. We also find that the
error correction coefficient in the Comp equation is highly significant and has the
correct (negative) sign, which suggests that the cointegrating relationship identified in
the previous cointegration tests for Singapore may be this Comp equation.
For Thailand, we find that none of the explanatory variables included in the TFP
equation Granger-causes TFP, but the error correction terms have the correct sign and at
least one of them is highly significant. We find that TFP Granger-causes Tradetax and
90
This negative causation implies that past values of Comp affect the current value of Human and that
Comp reduces the value of Human.
160
Comp, in the Tradetax equation and Comp equation respectively. Causal relationships
among the other variables under study are also found. For instance, we find that the two
types of government expenditures, Noncomp and Comp Granger-cause our openness
variable, Tradetax, and that Noncomp is a positive influence, while Comp is a negative
influence on Tradetax. Evidence of a significant bi-directional causality between Comp
and Human is also found.
The diagnostic tests on each of the ECM estimations indicate that some of the
regression equations suffer from mis-specification based on Ramseys RESET test.91
Unfortunately, the RESET test does not specifically identify what the problem with the
specification is. We also find that the residuals in some of the regression equations are
not normally distributed; hence the statistical significance of the variables as well as the
standard inferences from the t-statistics and F-statistics in these particular equations
may be less reliable than in the other equations.92 In Chapter 3, when we use crosssectional data, we deal with the problem of having non-normal residuals by first
identifying outliers and/or influential observations that may have affected the residuals
based on the RSTUDENT and Hat-values, and then re-estimating the equations without
the influential observations.
However, for time-series data, especially with lagged
terms in the equations, dealing with outliers and influential observations is not as
manageable, as dropping a time period in the middle of the series causes difficulties.
For these reasons, the Granger causality results presented above are at best tentative.
So far, we have looked at whether or not a specific variable Granger-causes TFP

(or another variable, i.e. Tradetax, Domtax, Noncomp, Comp, or Human, depending on
the specific equation analysed), holding the other variables constant. In Table 4.5, we
look at whether Tradetax, Domtax, Noncomp, Comp, and Human, jointly Granger-cause
TFP.
A standard F-test is used to determine whether the coefficients of all the
explanatory variables are equal to zero. We also test the significance of the explanatory
variables in the other ECM equations in each HPAE.
91
The equations that appear mis-specified based on Ramseys RESET test are: equation 4 of Appendix
L.2; and equations 2 and 3 of Appendix L.3.
92
The equations that appear to have non-normal residuals are: equations 1, 4 and 5 of Appendix L.1; and
equations 1 and 3 of Appendix L.2.
161
Table 4.5
Granger Causality Test Results

(H0: no causality)
Direction of causality
F-statistics
Conclusions
Indonesia:
Tradetax, Domtax, Noncomp, Comp, Human x TFP
TFP, Domtax, Noncomp, Comp, Human x Tradetax
TFP, Tradetax, Noncomp, Comp, Human x Domtax
TFP, Tradetax, Domtax, Comp, Human x Noncomp
TFP, Tradetax, Domtax, Noncomp, Human x Comp
F(6,14) = 0.7195
F(6,14) = 2.0773
F(6,14) = 0.7107
F(6,14) = 0.5039
F(6,14) = 4.6317
Dont reject
Dont reject
Dont reject
Dont reject
Reject at 1%
Malaysia:
TFP, Tradetax, Noncomp, Comp, Human x Domtax
TFP, Tradetax, Domtax, Noncomp, Comp x Human
F(8,13) = 3.0445
F(8,13) = 2.8579
F(8,13) = 3.0232
F(8,13) = 5.9205
F(8,13) = 7.5995
Reject at 5%
Reject at 5%
Reject at 5%
Reject at 1%
Reject at 1%
Singapore:
F(7,14) = 6.4546 Reject at 1%

F(7,14) = 14.285 Reject at 1%
F(7,14) = 20.973 Reject at 1%
Thailand:
TFP, Tradetax, Domtax, Comp, Human x Noncomp
F(8,13) = 8.2511
F(8,13) = 5.1252
F(8,13) = 2.7108
F(8,13) = 4.5406
F(8,13) = 6.2082
Reject at 1%
Reject at 1%
Reject at 10%
Reject at 1%
Reject at 1%
From Table 4.5 we find that Tradetax, Domtax, Noncomp, Comp, and Human
jointly Granger-cause TFP in three of the four HPAEs (Malaysia, Singapore and
Thailand). We are led to conclude then, that, as hypothesised, these variables, as a
group, are significant determinants of TFP in Malaysia, Singapore and Thailand.
Analysing the other ECM equations in Indonesia, we find that TFP, Tradetax,
Domtax, Noncomp, and Human jointly Granger-cause Comp. For the ECM equations in
the other HPAEs, we find that any other combination of variables, including TFP,
jointly Granger-cause the dependent variable in the respective ECM equations. This
result also suggests that TFP (jointly) Granger-causes its own hypothesised
162
determinants. Overall, the Granger causality test results provide evidence of multidirectional causation among the different variables.
4.8
This chapter examines the determinants of TFP using time-series data for five of
the HPAEs: Indonesia, Malaysia, Singapore, South Korea and Thailand. We conduct
our analysis in the context of TFP levels and hypothesise long-run cointegrating
relationships between TFP and its perceived determinants: the degree of openness,
government intervention and human capital. We derive a new set of TFP levels for the
five HPAEs based on a levels accounting methodology that excludes human capital
from the underlying production function.
We find that our TFP levels (and growth
rates) differ from the TFP levels and growth rates derived in other empirical studies
primarily because of the manner in which TFP has been calculated, i.e., we use levels
accounting while other studies used growth accounting or regression analysis; also we
exclude human capital from the underlying production function while others included
human capital.
We also hypothesise that openness, measured in terms of the ratio of taxes on

international trade to GDP (Tradetax) influences economic growth, but essentially
through TFP; hence, we test what we refer to as the openness-led-growth (OLG) via
TFP hypothesis, instead of the more common export-led-growth (ELG) hypothesis. We
also acknowledge that other factors, particularly government intervention and human
capital, could influence the openness TFP relationship. Government intervention for
our purposes is measured in terms of government expenditure and revenue, while
human capital is proxied by an index of labour quality, taken from a later version of the
Collins and Bosworth (1996) dataset.93
Total
central
government
expenditures
are
disaggregated
into
non-
complementary and complementary expenditures because different types of government

93
As noted in section 4.5, caution regarding the interpretation of results relating to human capital must be
made due to the concerns about Collins and Bosworths (1996) derivation of annual estimates of their
human capital index.
163
expenditures are likely to have different effects on TFP and growth.
Non-
complementary (Noncomp) expenditure is defined as expenditures that do not directly

affect the productivity of the private sector (i.e., general public services, defence, and
social security and welfare). Complementary (Comp) expenditure on the other hand is
defined as expenditures that are somehow conducive to private investment (i.e., health,
education and economic services, including expenditures on research, roads, other
transport and communication services). Government revenues are also disaggregated
into domestic tax revenues (Domtax) and international tax revenues (Tradetax), which
are essentially taxes imposed on international trade (we also use this as our measure of
openness).
We follow a three-step procedure in conducting Granger causality tests for the

HPAEs.
We first conduct unit root and multivariate cointegration tests following
Johansens (1988) procedure and error correction representation of the models to

determine whether long-run equilibria exist among our variables of interest, before
conducting Granger causality tests. We find support for the presence of at least two
cointegrating relationships each for Malaysia and Thailand, and at least one
cointegrating relationship for Indonesia and Singapore. One advantage of the Johansen
procedure is that it tests for the number of cointegrating relationships. However, if
more than one cointegrating relationship is found, the Johansen procedure is unable to
actually identify what exactly the cointegrating relationships are. In such situations we
should be able to turn to economic theory and economic models to help identify the true
cointegrating relationships. However, since a sound TFP theory and model is still
lacking, our cointegration results are at best tentative.
Standard diagnostic tests on the ECM equations indicate that some equations
suffer from mis-specification and non-normal residuals problems. For these reasons,
inferences based on standard t-tests and F-tests for some of the equations are less
reliable, and, in general, our results are at best tentative.
With this in mind, we conclude with caution that human capital is a significant
positive determinant of TFP but only for Malaysia. We fail to find any support for our
164
openness-led-growth (OLG) via TFP hypothesis. We find that Noncomp is a significant
negative influence on Malaysias TFP. We also find some support for our initial
hypothesis that government intervention influences the degree of openness, at least for
Malaysia and Thailand. We find that Comp negatively influences openness, measured
in terms of Tradetax, in both countries. Human capital also appears to positively
influence openness in Malaysia, and Noncomp is found to positively influence openness
in Thailand.
As a group, Noncomp, Comp, Domtax, Tradetax, and Human, are
significant determinants of TFP levels for Malaysia, Singapore and Thailand. There is
also evidence that TFP (jointly) causes its own determinants.
For South Korea, where all the non-stationary variables have been differenced,
so that all can be treated as stationary, standard OLS regressions show that Comp is the
only significant positive influence on TFP. This result however cannot be considered
robust since the adjusted R-squared for the whole regression is very low (0.020).
This chapter and the previous one attempt to determine the main factors that
affect TFP levels, using data for a cross-section of countries, and also, using time series
data for five HPAEs. We find substantial differences in the results of these two studies.
For example, in the cross-country analysis, total government consumption expenditure
is determined to be a positive influence on TFP. In the time-series analysis, government
intervention, measured in terms of government non-complementary expenditure,
Noncomp, is a significant negative influence on TFP, but only for Malaysia. There is no
evidence that human capital has any significant effect on TFP in the cross-section study;
but, in the time series study, at least for Malaysia, human capital appears to be a
significant positive influence on TFP. Here we note again, that while extreme care has
been made to ensure the accuracy of our time-series results, because of the short span of
time-series data available, our results are at best tentative. In the next chapter, we
continue our analysis of TFP levels and attempt to further examine the results from our
cross-section and time-series studies, by using panel data. The advantages of using
panel data over pure cross-sectional or time-series data in the analysis of economic
growth are discussed fully in the next chapter.
165
~ Appendix H.1 ~
TFP levels in the HPAEs Indonesia

(using the latest Collins and Bosworth (1996) dataset)
(without human capital in the production function)
TFP levels
= 0.2
Indonesia
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
777388.18
805528.70
802570.60
762752.43
778296.35
765886.93
773121.24
762657.95
848879.58
887082.07
934878.43
970215.10
996083.84
1048995.39
1089397.60
1095464.25
1128668.30
1189259.71
1233796.70
1256337.81
1300245.19
1344538.11
1274464.32
1343055.76
1392009.71
1381340.20
1411388.11
1432224.07
1470822.94
1555944.30
1641321.33
1734891.33
1802826.83
1875005.91
1954272.63
2050908.01
TFP levels
= 0.33
TFP levels
= 0.4
TFP levels
= 0.5
TFP levels
= 0.6
760413.95
788613.87
782100.15
735957.89
753171.00
737833.60
744328.46
732593.83
831601.51
872445.14
920638.43
952471.28
970556.53
1017925.34
1047702.23
1036598.22
1057702.94
1107348.12
1136895.10
1141085.17
1163625.52
1186370.40
1092111.90
1144408.50
1180774.46
1156833.07
1173737.79
1182005.88
1206743.24
1274979.29
1340524.53
1411035.15
1457711.21
1508553.05
1563672.87
1632681.03
749083.68
777315.13
768479.31
718277.29
736550.68
719346.60
725364.34
712830.07
820053.74
862630.23
911079.32
940591.21
953572.67
997325.03
1020264.70
998386.71
1012012.63
1054895.04
1075386.58
1068813.59
1079018.93
1089583.17
983254.56
1026389.63
1055756.44
1025392.94
1035427.84
1037321.86
1054804.37
1113497.66
1168126.69
1226078.88
1261593.31
1301187.82
1343678.02
1398140.64
727569.30
755842.31
742712.31
685159.52
705327.44
684771.97
689919.89
675974.28
798093.24
843891.85
892805.83
917951.23
921447.76
958518.75
969037.94
928185.89
928877.58
960061.61
965325.94
941328.75
931942.68
923665.38
801947.71
830912.20
849603.57
811354.75
811746.37
805076.17
812313.81
856093.71
894204.94
933402.47
953029.74
976509.51
1001076.82
1034693.39
696143.31
724432.60
705308.11
637867.42
660523.03
635526.28
639488.46
623730.66
765935.14
816273.04
865814.67
884682.28
874823.82
902579.15
896279.97
831111.83
815742.80
832365.44
819648.33
776541.34
746388.88
719132.70
588877.50
603294.26
611315.92
569050.58
561396.40
548345.52
546805.40
574830.07
596484.78
617447.53
623076.56
632109.58
640897.98
655725.40
growth rate 60-95 :
2.81
2.21
1.80
1.01
-0.17
growth rate 72-95 :
3.19
2.29
1.68
0.50
-1.25
166
~ Appendix H.2 ~
TFP levels in the HPAEs Malaysia
TFP levels
= 0.2
Malaysia
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
TFP levels
= 0.33
TFP levels
= 0.4
TFP levels
= 0.5
TFP levels
= 0.6
5003.02
5190.10
5305.42
5488.82
5563.61
5793.59
6043.68
6029.70
6326.83
6414.37
6555.10
6598.97
6970.00
7535.44
7784.71
7412.63
8038.13
8322.18
8519.96
8975.92
9200.10
9411.69
9546.45
9748.33
10141.33
9538.19
9246.33
9478.72
10115.91
10772.85
11508.75
12130.96
12659.91
13285.36
14061.23
14884.54
4827.23
4966.54
5016.35
5145.41
5161.48
5346.55
5550.94
5472.54
5739.93
5786.11
5864.85
5833.92
6141.37
6639.76
6780.44
6314.65
6881.93
7086.86
7203.55
7562.32
7656.17
7712.21
7701.56
7762.88
8020.06
7377.72
7076.73
7270.46
7825.95
8365.57
8950.72
9397.49
9754.80
10189.13
10727.36
11278.58
4711.24
4820.05
4828.84
4924.25
4904.76
5062.44
5239.02
5123.37
5372.26
5394.42
5437.48
5365.01
5634.92
6092.31
6172.60
5662.47
6192.24
6353.33
6426.59
6730.49
6757.11
6735.45
6655.18
6649.15
6837.13
6195.45
5900.10
6070.63
6572.62
7043.81
7544.36
7899.71
8170.23
8507.00
8924.26
9339.58
4493.98
4547.90
4484.62
4521.71
4442.31
4553.30
4682.69
4507.91
4724.41
4708.16
4694.75
4559.64
4767.95
5155.12
5143.84
4582.51
5044.53
5139.11
5149.49
5367.93
5302.13
5178.47
5012.48
4922.75
5015.83
4414.03
4145.32
4277.46
4683.83
5044.58
5414.00
5639.56
5791.54
5993.28
6243.56
6475.79
4183.73
4164.54
4009.24
3973.60
3823.35
3877.74
3950.24
3713.31
3888.56
3831.03
3758.10
3563.69
3701.67
4002.45
3902.28
3325.51
3697.70
3726.67
3681.13
3810.30
3671.88
3477.16
3262.32
3121.72
3136.95
2640.88
2428.20
2516.57
2803.27
3041.97
3274.75
3384.37
3438.50
3525.27
3633.89
3718.18
growth rate 60-95 :
3.16
2.45
1.97
1.05
-0.34
growth rate 72-95 :
3.35
2.68
2.22
1.34
0.02
167
~ Appendix H.3 ~
TFP levels in the HPAEs Singapore
TFP levels
= 0.2
Singapore
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
TFP levels
= 0.33
TFP levels
= 0.4
TFP levels
= 0.5
TFP levels
= 0.6
11554.95
11983.47
12205.65
12857.16
11220.66
11307.85
12054.01
13179.82
14467.40
15637.15
16777.88
17528.96
18544.82
19416.65
19306.75
18726.05
18912.78
19302.17
19868.53
20599.04
21397.59
22737.10
23366.07
24395.64
25567.87
23892.68
23618.95
25556.09
28272.21
30489.78
32622.94
34374.74
35732.84
38924.61
42514.59
45471.67
12336.43
12616.54
12609.70
13090.45
10828.15
10660.50
11274.12
12292.39
13397.53
14281.79
15023.36
15349.88
15965.88
16470.58
16003.08
15190.68
15171.89
15384.64
15747.92
16252.05
16774.48
17691.22
17889.25
18452.93
19142.53
17421.01
17043.51
18608.41
20872.48
22667.06
24367.27
25656.67
26551.74
29043.61
31872.56
34104.07
12897.81
13066.02
12892.08
13251.51
10569.12
10241.61
10772.82
11723.32
12715.57
13427.88
13936.30
14024.97
14420.24
14726.94
14085.74
13176.08
13060.29
13185.44
13445.66
13832.78
14215.50
14916.04
14918.21
15262.21
15722.69
14053.45
13652.18
14997.39
16980.87
18528.44
19982.04
21028.76
21696.53
23799.76
26202.01
28044.69
14061.51
13984.78
13458.39
13569.81
10083.85
9474.86
9862.45
10692.77
11489.29
11913.31
12045.56
11770.70
11834.04
11851.63
10994.93
9996.30
9763.56
9773.50
9893.23
10116.50
10309.02
10710.35
10485.89
10557.75
10730.24
9261.69
8874.75
9865.96
11376.35
12527.85
13594.85
14292.96
14660.33
16169.75
17913.27
19183.90
16027.82
15501.40
14364.19
14066.73
9390.71
8421.09
8627.54
9301.31
9852.84
9937.62
9657.99
9026.10
8771.49
8528.01
7554.07
6578.11
6283.14
6208.87
6215.17
6297.27
6335.57
6484.09
6146.37
6040.56
6014.77
4923.87
4621.18
5230.82
6200.57
6924.02
7584.76
7962.40
8094.59
9002.35
10067.76
10791.14
growth rate 60-95 :
3.99
2.95
2.24
0.89
-1.12
growth rate 72-95 :
3.98
3.35
2.93
2.12
0.90
168
~ Appendix H.4 ~
TFP levels in the HPAEs South Korea
TFP levels
= 0.2
South Korea
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1513905.11
1546383.02
1522417.12
1601752.28
1706686.06
1741738.12
1873516.91
1889232.93
2006436.38
2183440.99
2265323.76
2393081.88
2420286.34
2712062.87
2830366.96
2918995.81
3209074.69
3414854.46
3608014.34
3679154.77
3324816.94
3431322.26
3506728.53
3838703.60
4060536.25
4225184.57
4625464.54
5058334.70
5499299.77
5646233.48
5991259.64
6352779.57
6448287.13
6618484.41
7017376.96
7469148.34
TFP levels
= 0.33
TFP levels
= 0.4
TFP levels
= 0.5
TFP levels
= 0.6
1598654.22
1634155.64
1592083.49
1673295.53
1792876.87
1817159.30
1940813.58
1914108.08
1996423.87
2140591.52
2182411.51
2283473.80
2276209.42
2558227.58
2639403.25
2689457.15
2953960.21
3107916.31
3225652.24
3219664.91
2804118.99
2871835.78
2906434.98
3188479.71
3356895.88
3470389.42
3812537.78
4176484.14
4540015.79
4602596.90
4837190.65
5077023.90
5077128.37
5153521.24
5437438.50
5762403.17
1658977.66
1696669.82
1641268.82
1723761.66
1853959.76
1870302.70
1987950.85
1931210.01
1989643.92
2111935.23
2127771.42
2211822.11
2183167.89
2458634.58
2516961.59
2543769.85
2792167.81
2915112.73
2989062.94
2940446.31
2497445.12
2544452.26
2558069.67
2810444.61
2949435.59
3035710.32
3342982.79
3666376.53
3985174.50
4005428.04
4182194.71
4359213.04
4315352.86
4347294.12
4571551.88
4830474.33
1782650.89
1824929.58
1741125.97
1826111.34
1978536.69
1977940.81
2082745.64
1964845.29
1976548.51
2057398.82
2025579.86
2079088.52
2013275.23
2276239.87
2295259.26
2283110.56
2502962.09
2574326.03
2578197.14
2465675.44
1994590.92
2011665.62
1996537.83
2199792.98
2294282.17
2341225.58
2590190.80
2847196.14
3094263.56
3058377.97
3153140.57
3242649.82
3147499.71
3124604.90
3264626.19
3429799.65
1987824.62
2037974.78
1904121.04
1992883.25
2183422.34
2152948.21
2235057.22
2016927.39
1956871.01
1977439.35
1880018.38
1892997.01
1780713.64
2025335.03
1995987.19
1938158.36
2120820.46
2132346.52
2060699.67
1888257.17
1418767.53
1409132.99
1371495.78
1517679.14
1568035.13
1579458.83
1759743.45
1940992.85
2108906.09
2032260.48
2055386.31
2071051.48
1951251.52
1894463.09
1960074.43
2041427.04
growth rate 60-95 :
4.67
3.73
3.10
1.89
0.08
growth rate 72-95 :
5.02
4.12
3.51
2.34
0.60
169
~ Appendix H.5 ~
TFP levels in the HPAEs Thailand
TFP levels
= 0.2
Thailand
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
TFP levels
= 0.33
TFP levels
= 0.4
TFP levels
= 0.5
TFP levels
= 0.6
14885.79
15086.92
15760.28
16456.70
16837.88
17529.43
19030.22
19678.48
20507.51
21271.80
22458.94
22386.91
22198.72
23516.37
23440.81
23563.16
24947.06
26394.49
28210.77
28334.60
28381.48
29106.91
29673.57
30307.16
31050.93
31510.28
32375.72
34758.50
38755.19
42519.77
46173.65
48621.47
51340.88
54379.06
57887.07
61912.71
14572.35
14581.89
15103.94
15609.20
15700.17
16149.10
17426.85
17696.16
18179.25
18585.28
19392.92
18997.06
18549.96
19611.99
19324.96
19278.41
20445.81
21596.06
23087.06
22942.16
22754.43
23192.21
23515.79
23863.39
24319.28
24561.04
25212.54
27222.04
30668.89
33789.73
36607.76
38136.54
39924.93
41965.23
44368.82
47158.78
14362.99
14248.15
14673.31
15057.97
14970.75
15273.10
16414.43
16463.50
16748.93
16954.98
17550.76
16990.16
16417.74
17334.23
16947.19
16819.02
17858.38
18841.70
20145.46
19874.22
19579.75
19872.73
20075.79
20283.33
20596.17
20733.23
21269.96
23053.95
26157.03
28901.24
31261.86
32330.34
33649.23
35185.28
37028.43
39189.96
13965.12
13621.96
13872.12
14042.88
13650.16
13705.96
14614.00
14310.46
14285.73
14187.27
14459.43
13679.53
12953.08
13640.31
13134.41
12904.65
13733.27
14457.63
15462.37
15040.73
14626.08
14724.29
14768.59
14794.69
14917.73
14922.30
15290.12
16697.07
19205.89
21338.91
23010.77
23462.15
24143.82
24992.26
26065.76
27360.74
13383.20
12725.23
12740.76
12633.74
11867.83
11632.44
12255.16
11572.52
11226.09
10830.01
10781.00
9850.45
9044.89
9486.96
8926.93
8638.16
9224.52
9678.76
10355.80
9860.60
9401.34
9348.17
9275.10
9172.33
9150.72
9066.19
9272.68
10241.41
12027.38
13475.96
14464.02
14434.25
14600.50
14884.06
15313.09
15874.25
growth rate 60-95 :
4.16
3.41
2.91
1.94
0.49
growth rate 72-95 :
4.56
4.14
3.86
3.30
2.48
170
~ Appendix I.1 ~
Unit root test results for Indonesia
Country
Indonesia
Null hypothesis
Levels
First Difference
ln(TFP)
ln(TFP)
0.450 (0)
0.800 (0)
1.030 (0)
4.624 (0) *
7.243 (0) *
10.864 (0) *
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
No constant, no trend
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
1.484 (3)
1.226 (3)
ln(Tradetax)
ln(Tradetax)
2.384 (0)
2.561 (0)
2.907 (0)
4.252 (0) *
6.150 (0) *
9.149 (0) *
1.707 (0)
2.321 (0)
ln(Domtax)
ln(Domtax)
2.399 (5)
1.983 (5)
2.891 (5)
4.851 (0) *
7.858 (0) *
11.779 (0) *
2.128 (5)
2.345 (5)
ln(Noncomp)
ln(Noncomp)
2.032 (0)
5.219 (0) *
7.722 (0) *
5.612 (0) *
10.988 (0) *
16.332 (0) *
0.672 (0)
0.289 (0)
171
~ Appendix I.1 ~
(continued)
Country
Indonesia
Null hypothesis
Levels
First Difference
ln(Comp)
ln(Comp)
2.262 (0)
3.430 (0)
5.040 (0)
4.359 (0) *
6.690 (0) *
9.804 (0) *
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
0.968 (0)
0.544 (0)
ln(Human)
ln(Human)
0.977 (5)
15.044 (5) *
6.690 (5) *
6.664 (4) *
15.532 (4) *
23.207 (4) *
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Notes:
Numbers in parentheses indicate number of lags.
* rejects the null hypothesis at the 10% significance level based on McKinnon asymptotic
(1991) critical values.
172
~ Appendix I.2 ~
Unit root test results for Malaysia
Country
Malaysia
Null hypothesis
Levels
First Difference
ln(TFP)
ln(TFP)
0.851 (0)
0.975 (0)
0.877 (0)
4.034 (0) *
5.552 (0) *
8.316 (0) *
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
0.444 (0)
0.669 (0)
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
ln(Tradetax)
ln(Tradetax)
0.957 (0)
9.564 (0) *
5.615 (0) *
4.571 (0) *
7.180 (0) *
10.569 (0) *
2.483 (0)
10.677 (0)
ln(Domtax)
ln(Domtax)
2.600 (0)
2.660 (0)
3.476 (0)
3.427 (3) *
3.963 (3)
5.914 (3)
1.690 (4)
2.363 (4)
3.498 (3) *
6.150 (3) *
ln(Noncomp)
ln(Noncomp)
3.401 (1) *
4.064 (1) *
6.088 (1) *
4.602 (1) *
7.071 (1) *
10.605 (1) *
173
~ Appendix I.2 ~
(continued)
Country
Malaysia
Null hypothesis
Levels
First Difference
ln(Comp)
ln(Comp)
2.261 (0)
1.928 (0)
2.737 (0)
5.023 (1) *
8.459 (1) *
12.664 (1) *
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
1.782 (0)
1.734 (0)
ln(Human)
ln(Human)
1.615 (1)
3.094 (1)
4.366 (1)
2.420 (0)
2.301 (0)
3.169 (0)
2.606 (1)
3.661 (1)
0.749 (0)
0.510 (0)
0.792 (0)
Notes:
174
~ Appendix I.3 ~
Unit root test results for Singapore
Country
Singapore
Null hypothesis
Levels
First Difference
ln(TFP)
ln(TFP)
1.736 (1)
2.806 (1)
3.777 (1)
3.803 (1) *
4.998 (1) *
7.307 (1) *
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
0.361 (1)
0.389 (1)
ln(Tradetax)
ln(Tradetax)
3.577 (1) *
8.067 (1) *
6.398 (1) *
3.840 (2) *
4.982 (2) *
7.376 (2) *
ln(Domtax)
ln(Domtax)
2.878 (1)
2.975 (1)
4.362 (1)
3.378 (0) *
3.845 (0) *
5.741 (0) *
3.019 (1) *
4.663 (1) *
ln(Noncomp)
ln(Noncomp)
2.919 (0)
2.676 (0)
3.993 (0)
5.092 (0) *
8.733 (0) *
12.973 (0) *
2.878 (0) *
4.162 (0) *
175
~ Appendix I.3 ~
(continued)
Country
Singapore
Null hypothesis
Levels
First Difference
ln(Comp)
ln(Comp)
0.694 (0)
1.437 (0)
2.154 (0)
5.058 (0) *
8.559 (0) *
12.795 (0) *
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
1.370 (0)
0.940 (0)
ln(Human)
ln(Human)
0.649 (5)
2.880 (5)
4.264 (5)
3.371 (4) *
3.943 (4)
5.817 (4) *
1.803 (1)
1.661 (1)
Notes:
176
~ Appendix I.4 ~
Unit root test results for South Korea
Country
South Korea
Null hypothesis
Levels
First Difference
ln(TFP)
ln(TFP)
3.280 (5) *
4.895 (5) *
6.893 (5) *
3.622 (0) *
4.569 (0) *
6.826 (0) *
ln(Tradetax)
ln(Tradetax)
2.715 (1)
3.325 (1)
4.918 (1)
2.871 (4)
3.044 (4)
4.498 (4)
3.021 (1) *
4.634 (1) *
2.842 (4) *
4.108 (4) *
ln(Domtax)
ln(Domtax)
1.708 (0)
2.290 (0)
1.482 (0)
4.507 (0) *
7.051 (0) *
10.437 (0) *
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
0.765 (0)
2.129 (0)
ln(Noncomp)
ln(Noncomp)
2.031 (0)
1.415 (0)
2.078 (0)
6.331(0) *
14.101 (0) *
20.960 (0) *
1.868 (0)
1.790 (0)
177
~ Appendix I.4 ~
(continued)
Country
South Korea
Null hypothesis
Levels
First Difference
ln(Comp)
ln(Comp)
5.852 (0) *
12.491 (0) *
18.734 (0) *
5.582 (2) *
10.463 (2) *
15.623 (2) *
ln(Human)
ln(Human)
3.151 (1) *
4.744 (1)
4.966 (1) *
1.661 (0)
1.688 (0)
2.060 (0)
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
2.082 (0)
2.664 (0)
1.430 (0)
Notes:
178
Appendix I.5
Unit root test results for Thailand
Country
Thailand
Null hypothesis
Levels
First Difference
ln(TFP)
ln(TFP)
2.250 (1)
3.160 (1)
3.572 (1)
3.080 (5)
3.162 (5)
4.743 (5)
0.246 (1)
0.916 (1)
3.033 (1) *
4.764 (1) *
ln(Tradetax)
ln(Tradetax)
2.691 (0)
4.175 (0) *
3.637 (0)
4.748 (1) *
7.655 (1) *
11.302 (1) *
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
0.775 (0)
2.380 (0)
ln(Domtax)
ln(Domtax)
3.324 (3) *
6.901 (3) *
5.937 (3) *
3.477 (3) *
4.058 (3) *
6.087 (3) *
ln(Noncomp)
ln(Noncomp)
2.254 (1)
2.716 (1)
4.074 (1)
4.067 (0) *
6.333 (0) *
9.345 (0) *
1.553 (0)
1.206 (0)
179
~ Appendix I.5 ~
(continued)
Country
Thailand
Null hypothesis
Levels
First Difference
ln(Comp)
ln(Comp)
3.070 (1)
4.076 (1) *
5.330 (1)
4.131 (1) *
5.817 (1) *
8.667 (1) *
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
Constant, trend
=0
, , , = , 0, 0
, , = 0, 0, 0
Constant, no trend
=0
, , = 0, 0
=0
2.244 (4)
2.600 (4)
ln(Human)
ln(Human)
2.984 (5)
5.038 (5) *
6.240 (5) *
2.121 (0)
2.502 (0)
3.469 (0)
1.065 (1)
3.994 (1) *
2.523 (0)
3.470 (0)
0.322 (0)
Notes:
180
~ Appendix J ~
Plots of Human
for Malaysia, South Korea and Thailand
1.
Malaysia
Change in Human
1.5
1
0.5
2.
93
19
91
89
19
87
19
19
19
85
83
81
19
19
19
79
77
75
19
19
19
ye
ar
73
South Korea
Change in Human
4
3
2
1
93
19
91
19
89
19
87
85
19
83
19
81
19
79
19
77
19
75
19
19
Thailand
Change in Human
1.5
1
0.5
ar
19
73
19
75
19
77
19
79
19
81
19
83
19
85
19
87
19
89
19
91
19
93
0
-0.5
ye
3.
19
ye
ar
73
181
~ Appendix K.1 ~
Results of applying Johansens (1988) procedure for testing cointegration
for Indonesia

I(1)
I(1)
I(1)
I(1)
I(1)
I(0)
1. Choosing the appropriate lag length, k, using AIC, SBC and LR test
k
AIC
2
1
0
218.96
192.06
99.04
Constant, no trend
SBC
LR test
178.20
170.12
95.91
(1v2) 125.76
(0v1) 258.03
AIC
241.60
202.62
156.90
Constant, trend
SBC
LR test
197.74
177.56
150.64
(1v2) 149.96
(0v1) 313.40
Notes:
Because of the short span of time series data available, the maximum order for the ECM
possible is only 2.
For the Akaike information and Schwarz Bayesian criteria (AIC and SBC respectively), we
choose the lag length that maximises the AIC and SBC statistics, which is k = 2.
For the Likelihood ratio test (LR test), the hypothesis that k = 1 against the alternative that k = 2
is rejected, so we conclude that k = 2.
Based on the AIC, SBC and LR test we conclude that the appropriate lag length to use is k = 2.
182
~ Appendix K.1 ~
(continued)
2. -max and LR trace tests for model selection and determining the value of r
-max test
Null
r=0
Alternative
r=1
r1
r=2
r2
r=3
r3
r=4
r4
r=5
LR trace test
Null
r=0
Alternative
r1
r1
r2
r2
r3
r3
r4
r4
r5
Model 2
Model 3
Model 4
64.64
(34.40)
18.75
(28.27)
14.77
(22.04)
11.68
(15.87)
7.33
(9.16)
64.46
(33.64)
17.15
(27.42)
84.62
(37.86)
30.08
(31.79)
11.89
(21.12)
7.62
(14.88)
0.40
(8.07)
17.15
(25.42)
9.56
(19.22)
1.30
(12.39)
117.18
(75.98)
101.52
(70.49)
37.06
(48.88)
142.71
(87.17)
58.09
(63.00)
19.92
(31.54)
8.03
(17.86)
0.40
(8.07)
28.00
(42.34)
10.86
(25.77)
1.30
(12.39)
52.53
(53.48)
33.78
(34.87)
19.01
(20.18)
7.33
(9.16)
Notes:
Numbers in parentheses give 95% critical value (Johansen and Juselius, 1990).
indicates rejection of the null at 5% significance level,

indicates fail to reject the null.
Based on both -max and LR trace tests, Model 2 with r = 1 is chosen for Indonesia.
183
~ Appendix K.2 ~
for Malaysia

I(1)
I(1)
I(1)
I(0)
I(1)
I(1)
k
AIC
2
1
0
303.97
274.43
134.68
Constant, no trend
SBC
LR test
261.42
251.52
131.41
(1v2) 131.08
(0v1) 351.49
AIC
318.90
273.16
216.32
Constant, trend
SBC
LR test
273.07
246.98
209.78
(1v2) 163.48
(0v1) 185.68
Notes:
Because of the short span of time series data available, the maximum possible order for the
ECM is only 2.
184
~ Appendix K.2 ~
(continued)
Model 2
Model 3
Model 4
71.07
(34.40)
44.91
(28.27)
68.78
(33.64)
39.84
(27.42)
21.03 3
(21.12)
69.82
(37.86)
47.57
(31.79)
21.08
(25.42)
15.81
(14.88)
1.65
(8.07)
19.75
(19.22)
1.99
(12.39)
161.22
(75.98)
90.15
(53.48)
147.11
(70.49)
78.33
(48.88)
160.21
(87.17)
90.39
(63.00)
42.81
(42.34)
21.74
(25.77)
-max test
Null
r=0
Alternative
r=1
r1
r=2
r2
r=3
r3
r=4
r4
r=5
LR trace test
Null
r=0
Alternative
r1
r1
r2
r2
r3
45.24
(34.87)
38.49
(34.51)
r3
r4
r4
r=5
23.10
(20.18)
7.28
(9.16)
17.46
(17.86)
1.65
(8.07)
22.14
(22.04)
15.82
(15.87)
7.28
(9.16)
1.99
(12.39)
Notes:

Based on the -max test, Model 3 with r = 2 is chosen, but based on the LR trace test, Model 3
with r = 3 is chosen.
As mentioned in the text (section 4.6) the -max test is better than the LR trace test when the
power of the tests is low (Toda, 1995). Since we also acknowledge that our sample size is less
than desirable and that a low-order VAR model is used, we choose Model 3 with r = 2 for
Malaysia.
185
~ Appendix K.3 ~
for Singapore

I(1)
I(0)
I(0)
I(0)
I(1)
I(1)
k
AIC
2
1
0
246.79
226.80
151.94
Constant, no trend
SBC
LR test
204.24
203.88
148.67
(1v2) 111.98
(0v1) 221.70
AIC
272.37
241.39
189.94
Constant, trend
SBC
LR test
226.55
215.20
183.39
(1v2) 133.97
(0v1) 174.89
Notes:
ECM is only 2.
186
~ Appendix K.3 ~
(continued)
-max test
Null
r=0
Alternative
r=1
r1
r=2
r2
r=3
LR trace test
Null
r=0
Alternative
r1
r1
r2
r2
r=3
Model 2
Model 3
Model 4
29.67
(22.04)
27.52
(21.12)
37.05
(25.42)
17.93
(15.87)
15.98
(9.16)
17.93
(14.88)
2.96
(8.07)
18.58
(19.22)
5.01
(12.39)
63.59
(34.87)
48.41
(31.54)
60.64
(42.34)
33.92
(20.18)
15.98
(9.16)
20.89
(17.86)
2.96
(8.07)
23.58
(25.77)
5.01
(12.39)
Notes:

Based on both -max and LR trace tests, Model 4 with r = 1 is chosen for Singapore.
187
~ Appendix K.4 ~
for Thailand

I(1)
I(1)
I(0)
I(1)
I(1)
I(1)
k
AIC
2
1
0
333.69
286.58
163.60
Constant, no trend
SBC
LR test
291.14
263.67
160.32
(1v2) 166.21
(0v1) 317.98
AIC
355.70
303.87
241.65
Constant, trend
SBC
LR test
309.86
277.68
235.10
(1v2) 175.67
(0v1) 196.43
Notes:
ECM is only 2.
188
~ Appendix K.4 ~
(continued)
Model 2
Model 3
Model 4
47.09
(34.40)
46.96
(33.64)
49.00
(37.86)
45.84
(31.79)
30.06
(25.42)
-max test
Null
r=0
Alternative
r=1
r1
r=2
33.98
(28.27)
32.24
(27.42)
r2
r=3
r3
r=4
r4
r=5
26.45
(22.04)
20.07
(15.87)
0.45
(9.16)
20.78
(21.12)
14.77
(14.88)
0.18
(8.07)
LR trace test
Null
r=0
Alternative
r1
128.03
(75.98)
114.94
(70.49)
159.88
(87.17)
r1
r2
80.94
(53.48)
67.98
(48.88)
110.88
(63.00)
r2
r3
46.96
(34.87)
35.74
(31.54)
r3
r4
r4
r=5
20.51
(20.18)
0.45
(9.16)
14.95
(17.86)
0.18
(8.07)
65.03
(42.34)
34.97
(25.77)
20.64
(19.22)
14.34
(12.39)
14.34
(12.39)
Notes:

As mentioned in the text (section 4.6) the -max test is better than the LR trace test when the
power of the tests is low (Toda, 1995). Since we also acknowledge that our sample size is less
than desirable and that a low-order VAR model is used, we choose Model 3 with r = 2 for
Thailand.
189
~ Appendix L.1 ~
OLS estimations of the ECMs for Indonesia
1) dependent variable is lnTFPt (t-statistics are in parentheses):
lnTFPt
= 0.36lnTFPt-1 + 0.02lnTradetax t-1 0.18lnDomtax t-1

(1.18)
(0.53)
(0.18)
+ 0.17lnNoncomp t-1 + 0.05lnComp t-1 + 0.93lnHumant
(1.07)
(1.16)
(1.38)
0.6 t-1
(1.37)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
A:H0: residuals are uncorrelated
F(1,13) = 0.0089
Fail to reject
B:H0: model is correctly specified
F(1,13) = 0.0752
Fail to reject
C:H0: residuals are normally distributed
(2) = 29.0427
Reject at 1% level
F(1,19) = 0.4820 Fail to reject

D: H0: residuals are homoskedastic
Notes:
A: Lagrange multiplier test of residual correlation
B: Ramseys RESET test using the square of fitted values
C: Based on a test of skewness and kurtosis of residuals
D: Based on the regression of squared residuals on squared fitted values.
190
~ Appendix L.1 ~
(continued)
2) dependent variable is lnTradetaxt (t-statistics are in parentheses):

ln Tradetaxt = 0.01 ln Tradetax t 1 + 0.52 ln TFPt 1 1.04 ln Domtaxt 1
(0.053)
(0.352)
(1.28)
+ 0.48 ln Noncomp t 1 0.13 ln Comp t 1 0.54 ln Humant

(0.688)
(0.213)
(0.163)
0.041t 1
(0.162)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,13) = 1.0939
Fail to reject
F(1,13) = 0.0147
Fail to reject
2 (2) = 1.4040
Fail to reject

Notes:
191
~ Appendix L.1 ~
(continued)
3) dependent variable is lnDomtaxt (t-statistics are in parentheses):
ln Domtaxt = 0.17 Domtaxt 1 + 0.17 ln TFPt 1 + 0.01 ln Tradetax t 1
(0.414)
(0.231)
(0.103)
0.46 ln Noncomp t 1 + 0.26 ln Comp t 1 2.44 ln Humant

(1.299)
(0.848)
(1.451)
+ 0.161t 1
(1.452)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,13) = 0.2715
Fail to reject
F(1,13) = 1.7885
Fail to reject
2 (2) = 2.2313
Fail to reject

Notes:
192
~ Appendix L.1 ~
(continued)
4) dependent variable is lnNoncompt (t-statistics are in parentheses):
ln Noncomp t = 0.27 ln Noncomp t 1 0.75 ln TFPt 1 0.11 ln Tradetax t 1
(0.388)
(0.507)
(0.464)
+ 0.86 ln Domtax t 1 0.03 ln Comp t 1 1.25 ln Humant

(1.058)
(0.043)
(0.377)
+ 0.081t 1
(0.373)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,13) = 0.0370
Fail to reject
F(1,13) = 0.0605
Fail to reject
2 (2) = 46.9897
Reject at 1% level

Notes:
193
~ Appendix L.1 ~
(continued)
5) dependent variable is lnCompt (t-statistics are in parentheses):
ln Comp t = 0.26 ln Comp t 1 1.37 ln TFPt 1 0.02 ln Tradetax t 1
(0.861)
(1.857)
(0.146)
+ 0.77 ln Domtaxt 1 0.82 ln Noncomp t 1 5.89 ln Humant

(1.912)
(2.335)
(3.561)
+ 0.391t 1
(3.559)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,13) = 1.9581
Fail to reject
F(1,13) = 0.0548
Fail to reject
2 (2) = 17.8237
Reject at 1% level

Notes:
194
~ Appendix L.2 ~
OLS estimations of the ECMs for Malaysia

ln TFPt = 0.94 0.29 ln TFPt 1 0.14 ln Tradetax t 1 + 0.12 ln Domtaxt 1
(0.258) (0.729)
(0.779)
(0.962)
0.51 ln Comp t 1 + 69.90 ln Humant 1 0.35 ln Noncomp1

(0.509)
(3.286)
(2.179)
+ 0.06 1t 1 + 0.10 2t 1
(1.324)
(2.241)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,12) = 2.2391
Fail to reject
F(1,12) = 0.2201
Fail to reject
2 (2) = 5.0975
Reject at 10% level

Notes:
195
~ Appendix L.2 ~
(continued)
2) dependent variable is lnTradetaxt (t-statistics are in parentheses):
ln Tradetax t = 6.22 0.22 ln Tradetax t 1 + 0.62 ln TFPt 1 + 0.37 ln Domtaxt 1
(1.004) (0.743)
(0.919)
(1.700)
0.40 ln Comp t 1 + 125.65 ln Humant 1 + 0.01 ln Noncomp t

(2.462)
(2.312)
(0.031)
+ 0.011t 1 + 0.18 2t 1
(0.112)
(2.393)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,12) = 0.0108
Fail to reject
F(1,12) = 0.0051
Fail to reject
(2) = 1.6037
Fail to reject

Notes:
196
~ Appendix L.2 ~
(continued)
3) dependent variable is lnDomtaxt (t-statistics are in parentheses):

ln Domtaxt = 16.12 + 0.11Domtaxt 1 0.70 ln TFPt 1 + 0.42 ln Tradetax t 1
(2.317) (0.453)
(0.938)
(1.239)
0.28 ln Comp t 1 + 116.07 ln Humant 1 + 0.52 ln Noncomp t

(1.563)
(1.901)
(2.572)
+ 0.221t 1 0.03 2t 1
(2.698)
(0.317)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,12) = 0.0922
Fail to reject
F(1,12) = 0.9166
Fail to reject
2 (2) = 6.2091
Reject at 5% level

Notes:
197
~ Appendix L.2 ~
(continued)

ln Comp t = 31.71 0.12 ln Comp t 1 0.69 ln TFPt 1 + 0.06 ln Tradetax t 1
(4.136) (0.598)
(0.833)
(0.164)
+ 0.08 ln Domtaxt 1 + 85.65 ln Humant 1 + 0.78 ln Noncomp t

(0.281)
(1.273)
(3.545)
+ 0.261t 1 0.34 2t 1
(2.848)
(3.748)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,12) = 0.0000
Fail to reject
F(1,12) = 3.3755
Reject at 10% level
2 (2) = 0.3439
Fail to reject

Notes:
198
~ Appendix L.2 ~
(continued)
5) dependent variable is lnHumant (t-statistics are in parentheses):

ln Humant = 0.05 + 1.22 ln Humant 1 + 0.003 ln TFPt 1 0.003 ln Tradetax t 1
(1.522) (4.239)
(0.741)
(1.613)
+ 0.002 ln Domtaxt 1 0.0005 ln Comp t 1 0.002 ln Noncomp t

(0.642)
(1.562)
(2.352)
0.00031t 1 + 0.001 2t 1
(0.780)
(2.086)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,12) = 0.4934
Fail to reject
F(1,12) = 0.2685
Fail to reject
2 (2) = 1.6044
Fail to reject

Notes:
199
~ Appendix L.3 ~
OLS estimations of the ECMs for Singapore

ln TFPt = 30.52 + 0.03 ln TFPt 1 0.13 ln Comp t 1 7.39 ln Humant 1
(1.128)
(1.980) (0.125)
(1.547)
+ 0.06 ln Tradetax t 0.09 ln Domtaxt 0.08 ln Noncomp t

(1.358)
(0.645)
(0.891)
+ 0.081t 1
(1.977)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,13) = 0.3245
Fail to reject
F(1,13) = 1.0284
Fail to reject
2 (2) = 0.1594
Fail to reject

Notes:
200
~ Appendix L.3 ~
(continued)

ln Comp t = 150.27 + 0.17 ln Comp t 1 0.87 ln TFPt 1 17.92 ln Humant 1
(7.072)
(2.629)
(1.064)
(2.722)
0.25 ln Tradetaxt + 1.15 ln Domtaxt + 0.56 ln Noncomp t

(4.269)
(5.693)
(4.254)
0.42 1t 1
(7.156)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,13) = 0.2220
Fail to reject
F(1,13) = 0.6804
Fail to reject
2 (2) = 3.3316
Reject at 10% level

Notes:
201
~ Appendix L.3 ~
(continued)
3) dependent variable is lnHumant (t-statistics are in parentheses):

ln Humant = 1.04 + 0.40 ln Humant 1 + 0.02 ln TFPt 1 + 0.003 ln Comp t k
(1.802) (2.268)
(1.926)
(0.722)
+ 0.001 ln Tradetax t + 0.02 ln Domtaxt + 0.01 ln Noncomp t

(0.928)
(3.826)
(2.364)
0.00031t 1
(0.928)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,13) = 0.4033
Fail to reject
F(1,13) = 3.2468
Reject at 10% level
2 (2) = 0.3312
Fail to reject

Notes:
202
~ Appendix L.4 ~
OLS estimations of the ECMs for Thailand
1. dependent variable is lnTFPt (t-statistics are in parentheses):

ln TFPt = 0.96 + 0.51 ln TFPt 1 + 0.05 ln Tradetax t 1 0.05 ln Noncomp t 1
(0.400) (2.540)
(0.251)
(0.536)
+ 0.05 ln Comp t 1 6.427 ln Humant 1 0.03 ln Domtaxt

(1.700)
(1.212)
(0.271)
0.021t 1 0.05 2t 1
(0.990) (2.246)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,12) = 0.0001
Fail to reject
F(1,12) = 0.6176
Fail to reject
2 (2) = 0.2834
Fail to reject

Notes:
203
~ Appendix L.4 ~
(continued)
2. dependent variable is lnTradetaxt (t-statistics are in parentheses):

ln Tradetax t = 3.04 + 0.70 ln Tradetax t 1 + 1.09 ln TFPt 1 + 1.79 ln Noncomp t 1
(0.461) (2.624)
(1.981)
(3.256)
0.34 ln Comp t 1 8.42 ln Humant 1 0.08 ln Domtaxt

(4.546)
(0.577)
(0.290)
0.091t 1 0.24 2t 1
(1.387) (3.745)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,12) = 0.2480
Fail to reject
F(1,12) = 0.2486
Fail to reject
2 (2) = 1.4218
Fail to reject

Notes:
204
~ Appendix L.4 ~
(continued)
3. dependent variable is lnNoncompt (t-statistics are in parentheses):

ln Noncomp t = 7.45 0.10 ln Noncomp t 1 0.32 ln TFPt 1 0.13 ln Tradetax t 1
(1.478) (0.024)
(0.761)
(0.638)
+ 0.03 ln Comp t 1 + 23.25 ln Humant 1 0.42 ln Domtax t

(0.024)
(2.078)
(2.072)
0.041t 1 + 0.13 2t 1
(0.782)
(2.726)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,12) = 0.0047
Fail to reject
F(1,12) = 0.7706
Fail to reject
2 (2) = 0.0888
Fail to reject

Notes:
205
~ Appendix L.4 ~
(continued)
4. dependent variable is lnCompt (t-statistics are in parentheses):

ln Comp t = 19.76 + 0.06 ln Comp t 1 2.12 ln TFPt 1 0.18 ln Tradetax t 1
(2.898) (0.781)
(3.727)
(0.668)
0.11 ln Noncomp t 1 + 30.45 ln Humant 1 0.73 ln Domtaxt

(0.194)
(2.014)
(2.699)
0.151t 1 + 0.19 2t 1
(2.302)
(2.889)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,12) = 0.2180
Fail to reject
F(1,12) = 0.8448
Fail to reject
2 (2) = 0.5883
Fail to reject

Notes:
206
~ Appendix L.4 ~
(continued)
5. dependent variable is lnHumant (t-statistics are in parentheses):

ln Humant = 0.20 + 0.38 ln Humant 1 + 0.002 ln TFPt 1 0.01 ln Tradetax t 1
(1.338) (1.112)
(0.177)
(1.026)
0.01 ln Noncomp t 1 + 0.004 ln Comp t 1 + 0.005 ln Domtaxt

(1.084)
(2.143)
(0.866)
+ 0.002 1t 1 0.00 2t 1
(1.331)
(0.279)
Diagnostic tests:
Null hypothesis
Test statistics
Conclusion
F(1,12) = 0.2174
Fail to reject
F(1,12) = 1.2567
Fail to reject
2 (2) = 2.2577
Fail to reject

Notes:
207
~ CHAPTER FIVE ~
The determinants of Total Factor Productivity:
Evidence from a Panel of HPAEs
5.1
Introduction
In our study of the factors that are most likely to influence total factor
productivity (TFP) levels, we have so far used two types of data: one, a cross-section of
developed and developing countries, and the other, time-series data for five HPAEs:
Indonesia, Malaysia, Singapore, South Korea and Thailand. The results from using
these two data sets, however, have been very different. In the cross-section study, only
total government consumption expenditure has a statistically significant positive
influence on TFP. Human capital does not appear to have a significant influence. On
the other hand, the time-series analysis shows that government non-complementary
expenditure has a significant negative influence, at least for Malaysia and South Korea,
and that human capital in Malaysia is a significant positive influence on TFP.
In this chapter, we attempt to further examine the results from our cross-section
and time-series studies by using panel data and applying appropriate methods to deal
with non-stationarity in the panel data. By using panel data, we essentially combine
information from the time-series and cross-sectional dimensions of the data. We are
hopeful that inference about the presence of unit roots will be more accurate when the
cross-sectional dimension of the data is included, particularly in our case, because of the
low number of time-series observations. We also apply one of Pesaran et al.s (2001)
approaches to testing the existence of a level relationship between a dependent variable
and a set of regressors when it is uncertain whether the variables being considered are
stationary or non-stationary.
In this part of the study, we continue to hypothesise that openness (measured in

terms of the taxes imposed on international trade), the different roles of government,
and human capital are the major factors that drive TFP in the HPAEs. As in the
previous chapter, we include various government proxies in order to capture the
208
different roles of government that could have different effects on TFP. The main
contribution of this part of our study stems from the application of recent panel
econometric methods to test panel unit roots, specifically, Im et al.s (IPS) (1997) panel
unit root tests, and the application of a recently developed bounds testing procedure on
panel data (Pesaran et al. 2001).94,95 In our study, we use panel methods to test a TFP
equation in levels instead of a typical economic growth model.
This chapter is organised as follows: Section 5.2 highlights the advantages of

using panel data and provides a brief tabular review of growth studies utilising panel
data that include the HPAEs. Section 5.3 describes the panel data used in our analysis.
Section 5.4 reviews some of the recently developed panel unit root methods. Section
5.5 presents the panel unit root test results and also compares the individual country unit
root test results in this chapter with those obtained in Chapter 4. Section 5.6 details the
bounds testing strategy in the analysis of level relationships, and Section 5.7 concludes.
5.2
Panel data analysis
Much of the existing empirical work on the determinants of economic growth

relies on cross-country data. The basic problem with this approach however is that, in
general, cross-country regressions implicitly assume that the countries under study have
similar structural characteristics (e.g. different countries have the same production
functions) or that these structural differences are minor and do not significantly affect
economic growth (Miller, 1996). These regressions therefore fail to capture important
country-specific characteristics crucial to growth in individual countries, which could
have significant implications for the estimation results obtained.
94
While there have been several applications of panel data methods in recent years, many of these have
been on purchasing power parity (PPP) and exchange rate models (e.g., Oh, 1996; MacDonald, 1996;
Coakley and Fuertes, 1997; Pedroni, 1997), wages (e.g., Fleissig and Strauss, 1997; Breitung and Meyer,
1994), inflation rates (e.g., Culver and Papell, 1997), and growth models (e.g., Pedroni, 1996;
Demetriades and Hussein, 1996; McKoskey and Kao, 1998; see also Table 1), rather than on TFP.
95
We note that, Banerjee et al. (2001) warn against using panel methods (specifically with reference to
testing the validity of purchasing power parity (PPP)). Their Monte Carlo tests on panel unit root tests,
e.g. IPS (1997), and Levin and Lin (1992, 1993) show that panel unit root tests reject too often the null of
a unit root when there are cross-unit cointegrating relationships.
209
For instance, Mankiw et al. (1992) (hereafter MRW), use cross-country data to
test the basic predictions of the Solow (1956) growth model, which they augment with a
human capital variable, and find strong support for conditional convergence at about the
rate predicted by the original Solow model. On the other hand, Knight, et al. (1993)
and Islam (1995), using panel data sets consisting of the same set of countries as in
MRW across five-year time intervals over the period 1960-1985, find evidence of
higher rates of conditional convergence.
Moreover, how specific variables are found to affect growth also differs
substantially depending on whether panel or cross-sectional data are used in estimating
the relationships.
Again, MRW, using cross-sectional data, find a positive and
significant role for human capital in explaining growth, while Knight et al. and Islam,
find a significant but negative relationship between human capital and growth. Knight
et al. and Islam explain that the effect of time-series variation in the human capital data
is strong enough to outweigh the positive cross-sectional effects in the estimation.
However, considering the composition of the panel data they use, i.e. a larger number of
cross-sectional observations compared to the number of time-series observations, it is
unclear why the time-series properties of any variable outweigh the cross-sectional
properties. Mairesse (1990, p. 87) notes, the dispersion of changes over time in the
(level) variables is usually very small in comparison with the dispersion of differences
across individuals.
Similarly, when Harrison (1996) (see Chapter 2) estimates the relationship

between different measures of openness and economic growth using two data sets, one a
panel data set and the other a pure cross-sectional data set, she finds that the correlations
in the panel data are more statistically significant than the correlations obtained in the
pure cross-sectional data. Moreover, only one of the seven openness variables is found
to positively affect economic growth using the cross-sectional data, whereas six of
seven measures are found to positively affect growth using the panel data.
There are several advantages of using panel data over pure cross-sectional or
time-series data in the analysis of economic growth. First, the central focus in the study
210
of growth, which is long-run steady-state growth, is an analysis of what happens to the
variables of interest (e.g. GDP, labour, capital, etc.) over time; hence time-series data
have been argued to be essential in providing more dynamic information than would
otherwise be obtained when only cross-sectional data are used (Easterly, et al. 1997).
Other studies, however, e.g., Pesaran and Smith (1995), Phillips and Moon (1999) and
Temple (1999a) point out that pure cross-section regressions can produce consistent
estimates of the average long-run effects. For instance, Baltagi and Griffin (1983), in
their study of gasoline demand in the OECD using a panel data set, raised the possibility
that between estimates, which are computed from the cross-sectional dimensions of
the data, reflect long-run responses, while the within estimates, computed from the
time-series dimensions of the data, reflect short-run responses. The reason for this is
that each country has already adapted its own auto efficiency, utilisation and driving
infrastructure to long-standing patterns of price and income (p. 126); hence, the
variation between countries could more adequately represent long-run responses.
Secondly, estimations that make use of time-series data (especially if there are
not enough time-series observations) for a single group/country, are often subject to
collinearity problems. The use of panel data provides the researcher with a larger data
set and, consequently, additional information and increased degrees of freedom which
could be important in reducing collinearity among the explanatory variables. Moreover,
by making use of information from time-specific and country-specific effects, the
researcher is also better able to control for the effects of missing or unobserved
variables (Hsiao, 1986).
In spite of the obvious advantages of using panel data, it has only been over the
last decade that empirical studies of economic growth with panel data have appeared in
the literature. This is most likely due to the fact that comparable data for a wide range
of countries over several time periods have been difficult to obtain, and only with the
availability of the Summers and Heston (1991) data set has working on growth empirics
211
with panel data become more feasible. Table 5.1 presents a summary of some panel
data growth studies that have included the HPAEs in the panel.96
A critical aspect in the estimation of relationships that combine cross-sectional

and time-series data lies in the specification of the model, i.e., the model should be
specified in such a way that possible differences in behaviour across countries, as well
as any differences in behaviour over time for any given country, are adequately
captured.
In estimating panel data models, it is also important to examine the time-series

properties of the data, particularly with respect to the stationarity of the time-series data,
so it is desirable that the appropriate unit root and cointegration tests for panel data are
carried out. This is because many economic time-series have been found to exhibit the
characteristics of an integrated process of order one (I(1)), i.e., the series needs to be
differenced once to be stationary, and treating I(1) series in regressions as if they were
stationary (I(0)), can lead to spurious results.
Of the studies in the summary table, only Lee et al. (1996) attempt to examine
the order of integration of the variables and consequent cointegrating relationships in
panel data. This does not mean however, that this is the only panel growth study to do
so. For example, Demetriades et al. (1998) analyse endogenous growth using panel
data, giving due consideration to the time-series properties of the variables of interest.
Their panel dataset, however, do not include any of the HPAEs; hence their study is not
included in the summary table below.
96
Other important contributions to the literature on panel data growth studies include (with countries
examined in parentheses): Easterly et al. (1997) (Latin American countries); Evans (1998) (developed and
developing countries, but Thailand is the only HPAE in the panel); Evans and Karras (1996a, b) (US
states); Cellini (1997a) (US, France, Italy and Japan); Lee et al. (1998) (22 OECD countries);
Demetriades et al. (1998) (20 developed and developing countries).
212
Table 5.1
Selected panel data studies on growth
Author
Dataset
Methodology
Results
Knight,
Loayza
and
Villanueva
(1993)
98 non-oil
countries
(including Hong
Kong, Indonesia,
Malaysia, South
Korea,
Singapore and
Thailand) over
5-year time
intervals: 1965,
1970, 1975,
1980 and 1985
Chamberlains (1984)
estimation procedure, which
examines country-specific
effects bias, is used to test the
basic model specifications in
MRW, but including a proxy
for the economys trade
regime, and the level of
public investment
higher rates of convergence

are evident from the use of
panel data, relative to MRW
estimates from cross-section
data
Barro and
Lee
(1994)
116 countries
total, divided
into two: one
with 85 countries
with averaged
time-series
observations
from 1965-75;
and the other
with 95 countries
with averaged
time-series
observations
from 1975-85
SUR estimation to test the

relationship between real per
capita growth rates and two
sets of variables: initial levels
of state variables (initial per
capita GDP, initial stock of
educational human capital
and initial stock of human
capital); and control variables
(ratio of domestic investment
to GDP, ratio of government
consumption to GDP,
measure of trade distortion
(black market premium), and
measure of political
instability (average number of
revolutions))
a country with lower initial

level of real per capita GDP
grows faster (conditional
convergence effect) relative
to initial levels of human
capital; ratio of investments to
GDP positively affects
growth, while large ratio of
government consumption to
GDP, market distortions and
political instability negatively
affect growth
Loayza
(1994)
98 countries
(including Hong
Kong, Indonesia,
Malaysia,
Singapore, South
Korea, and
Thailand); 5-year
time intervals:
1965, 1970,
1975, 1980 and
1985
Chamberlains (1984)
estimation procedure is used
to estimate the rate of
convergence of an economy
to its own steady state based
on a standard neoclassical
growth model
estimated rate of convergence

is 0.0494 (higher relative to
MRW estimates, but similar
to that obtained by Knight et
al. (1993) and Islam (1995))
trade openness and level of

public investments
significantly and positively
affect growth
continued over page
213
Author
Dataset
Methodology
Results
Islam
(1995)
96 non-oil
countries
(including Hong
Kong, South
Korea, Malaysia,
Singapore and
Thailand); 75
inter countries
and 22 OECD
countries; 5-year
time intervals:
1965, 1970,
1975, 1980 and
1985
for the panel data, used

minimum distance (MD)
estimation procedure with
correlated effects, and leastsquares with dummy variables
(LSDV) estimation following
MRW specifications
regardless of estimation
technique used for panel data,
higher rates of convergence
were obtained relative to
MRW estimates
Taylor
(1995)
207 countries, 4
time periods:
1970-74;197579;198084;1985-89
2SLS (2 stage least squares)

growth accounting regression;
Collins
and
Bosworth
(1996)
88 countries
including 9 East
Asian countries
(Indonesia,
Korea, Malaysia,
Philippines,
Singapore,
Thailand,
Taiwan) for the
period 1960-94
growth accounting exercise,

which decomposed growth in
output per worker into the
contributions of physical and
human capital and a residual
measuring total factor
productivity growth
negative temporal relationship

between human capital and
growth is more evident in
panel context
high investment rates were the

key to high growth rates in the
Asia-Pacific region, which in
explanatory variables - growth turn are caused by low
rate of GDP per capita, initial distortions, among others;
per capita GDP, initial years
much of the investment in the
of primary and secondary
developing economies are in
schooling per person in the
the form of imported capital
population, 4 endogenous
goods, so high investment
control variables related to
environment might be more
factor accumulation, 10
aptly termed import-oriented
environmental variables
growth
that measure natural resource
endowments, demographic
structure, government
policies, price distortions and
openness, and financial
intermediations and monetary
instability
East Asian growth is due
largely to factor
accumulation, not TFP
growth
continued over page
214
Author
Dataset
Methodology
Results
Lee,
Pesaran
and Smith
(1996)
102 countries
with annual timeseries
observations
over the period
1960-89; and two
subsets, one
consisting of 61
developed and
less developed
countries, and
the other
consisting of 22
OECD countries
a stochastic Solow growth

model is developed to
examine the different
convergence hypotheses
steady-state growth rates

differ significantly across
countries; estimated speed of
convergence from a panel
data set is much higher (30%
per annum) relative to the
estimated speed of
convergence from crosssectional data (2% per annum)
several data sets

(panel and crosscountry)
depending on the
availability of
openness
measures for
each country
rank correlations across

different openness measures;
OLS regressions based on
pure cross-country data and
then on panel data; Granger
causality tests between
openness and growth
found positive relationships

among the seven openness
measures, but correlations
using panel data are more
statistically significant than
correlations using pure crosssectional data
explanatory variables - GDP

growth rate, capital stock,
years of primary and
secondary education,
population, labour force,
arable land and technological
change; openness measures trade reform (1960-84), trade
reform (1979-88), black
market premium, Dollars
price distortion index, trade
shares, disprotection of
agriculture
only one of seven openness

variables positively affects
growth using cross sectional
data; while six of seven
measures affect growth using
panel data
growth accounting exercise,

complemented by OLS (real
GDP growth rate on initial
labour productivity, growth of
the workforce, and average
investment ratio)
productivity growth in the

East Asian economies, in
general, has been a significant
factor in their rapid growth
since the 1950s
Harrison
(1996)
cross-country:
from 17 to 51
countries
time period:
annual from
1960-87 or 197888
Drysdale
and
Huang
(1997)
APEC member
countries,
including the
HPAEs, over
four sub-periods:
1950-59, 196069, 1970-79 and
1980-90
time-series properties of the

panel data are analysed based
on IPSs panel unit root tests
causality between openness

and growth was found to run
in both directions
capital accumulation has also

been an equally important
factor
continued over page
215
Author
Dataset
Methodology
Results
Cellini
(1997b)
84 countries
(including Hong
Kong, South
Korea, Malaysia
and Thailand);
annual timeseries data from
1960-85
real GDP per worker is

regressed on the share of
investment in GDP, secondary
school enrolment rate (proxy
for human capital as in
MRW), and the growth rate of
employment, using five
different estimators: total
OLS, between, within or
fixed effects, random
effects, and total IV
(instrumental variable)
based on the fixed-effect

estimator (which according to
the author is the most
appropriate), a significant
conditional convergence
effect is observed; country
fixed-effects, deriving from
different efficiency levels
across countries, are very
important in the growth
process; the effect of human
capital on growth is negative
once fixed effects are allowed
for (consistent with the results
obtained by Knight et al.
(1993) and Islam (1995))
GMM estimation procedure

on a basic neoclassical growth
model and an endogenous
growth model; ancillary
variables (measures of
political instability, income
distribution and financial
development) are introduced
into the basic growth models
ancillary variables affect

growth through factor
accumulation
sub-groups of G7, European,

African and
Asian countries
Benhabib
and
Spiegel
(1997)
5.3
92 developed and
developing
countries over 5year growth
periods from
1960-85
however, when country-fixed

effects are introduced in the
models, the effects of
ancillary variables on the rates
of factor accumulation are
insignificant or, when
significant, have the wrong
signs
Panel data
As in the previous chapters, we assume that TFP is a function of human capital,

openness and government-related variables, as these variables are considered as the
more controversial factors affecting economic growth in the HPAEs (see Chapter 2):
TFPi ,t = 0 + 1Oi ,t + 2, x Gi , x ,t + 3 H i ,t + i ,t
x
(5.1)
216
where,
TFPi,t
represents total factor productivity in country i and

time t
Oi,t
openness variable for country i and time t
Gi,x,t
government-related variable x for country i and time t
Hi,t
represents human capital in country i and time t
The TFP levels are obtained based on our improved levels accounting
equation.97
For the analysis in this chapter, we are unable to fully exploit the
advantages of panel characteristics because of the small size of our panel data, i.e., five
cross-sections and 18 time series observations, and we have to assume homogeneity
across our cross-section of countries and pool our data. In pooling the individual
country data available, our TFP levels need to be calculated from data using a common
currency, at least for the data on the level of output per worker (y/l).98 We take the y/l
data from Summers and Hestons (1991), Penn World Table 5.6. The other relevant
data for calculating the TFP levels are taken from a later version of the Collins and
Bosworth (1996) data set. We expect that the use of international prices to facilitate the
pooling of our data (as well as the difference in the number of time-series observations
used to keep the panel balanced) would affect the time-series properties not only of our
new TFP levels but also the rest of the variables of interest. We also use Collins and
Bosworths human capital data (Human), even though we have some reservations about
how they actually derive their human capital index.99 The complete data set for the
following estimations is presented in Appendix M.
97
Specific details on how the TFP levels are calculated, as well as how these TFP levels compare with
estimates from other studies are discussed in Chapter 4, Section 4.3.
98
In Chapter 3, Section 3.3.2, we discuss how conversion to international prices can distort the data by
imposing a set of international relative prices on each country (Knowles, 2001), with respect to the
government expenditure and capital-output ratio variables. However, panel analyses generally make use
of international prices (or a common currency), so in order to be compatible with these studies, we also
use international prices. Also, since the HPAEs are regionally connected and have more in common
among them than with other countries in the world, the problems associated with using international
problems may not be as great
99
See Chapter 4, Section 4.5.3 for a discussion of these reservations.
217
We include several measures of government in our analysis because we
hypothesise that different government actions could have different effects on TFP. In
many of the previous studies on growth and government, government expenditure has
always been used to proxy for the role of government, but this is actually a poor proxy
because even different types of government expenditures can affect the economy in
many ways.
For this reason, we disaggregate government expenditure into non-
complementary and complementary expenditures.
Government non-complementary
expenditures (Noncomp) are those that are considered as not directly affecting the
productivity of the private sector (e.g., general public services, defence, and social
security and welfare). Based on the results of the previous chapter we hypothesise that
this type of expenditure has a negative effect on TFP. On the other hand, government
complementary expenditures (Comp) are those that are likely to affect private
investment (e.g., health, education and economic services, including expenditures on
research, roads, other transportation and communication services). We hypothesise that
this type of expenditure has a positive effect on TFP.
We also include government revenues in our analysis and following the same
line of reasoning that different types of government revenue could affect the economy
in different ways, we also disaggregate our government revenue measures. In general,
government revenue in this study is defined as the taxes received by the central
government, and, depending on how these taxes are spent, may or not be positively
associated with TFP.
As mentioned in the previous chapter, we disaggregate
government revenue into domestic (Domtax) and international tax revenues (Tradetax),
instead of into distortionary and non-distortionary revenue as in Kneller et al. (1999)
and Bleaney et al. (2001), because we will be using Tradetax as our measure of
openness. Also, as in the previous chapter, we include only one measure of openness in
our TFP equation. This is because our correlation tests in Chapter 3 show that the
different openness measures are statistically significantly correlated, suggesting that
these variables measure or reflect the degree of openness similarly.
appears no need to include more than one measure of openness.100
100
See Chapter 3, Section 3.3.2 for some discussion on this point.
Hence, there
218
Data for our government and openness indicators are taken from various issues
of the IMFs International Financial Statistics (IFS) and Government Finance Statistics
(GFS) yearbooks, and the Asian Development Banks (ADB) Key Indicators for
Developing Member Countries. In order to keep the panel balanced, the full set of panel
data we use in this study consists of time-series observations over the period 1973 to
1990, a total of 18 annual observations for five HPAEs: Indonesia, Malaysia, Singapore,
South Korea and Thailand. The analysis is limited to these five HPAEs mainly because
the International Monetary Funds (IMF) Government Finance Statistics (GFS), the
main source of our government-related variables, does not report data for the other two
HPAEs: Hong Kong and Taiwan.
5.4
Panel unit roots
Over the past decade, several panel unit root tests have been put forward, the
most often cited ones being those of Levin and Lin (1993), Quah (1994), Im et al.
(1997) and Pedroni (1997). Baltagi and Kao (2000) present a survey of some of these
panel unit root and cointegration tests, noting the advantages and limitations of each
test.
In this study the panel unit root tests suggested by Im et al. (1997) (hereafter
IPS), are applied to test for non-stationarity of the relevant variables. Essentially, the
IPS panel unit root tests combine information on the stationarity or non-stationarity
characteristics of the time-series data for each country in the cross-section to give a
conclusion for the entire panel. IPS propose two panel unit root tests based on the mean
of individual unit root statistics: the LM-bar test and t-bar test. The LM-bar test is based
on the average of the Lagrange multiplier statistics computed for each group (in this
case, countries) of the panel, while the t-bar test is based on the average of DickeyFuller (DF) t-statistics computed for each group in the panel. The relevant DickeyFuller regression is expressed in terms of the following general model (with trend):101
y i ,t = i + i (t T 2) + i y i ,t 1 + i ,t , i = 1,..., N ; t = 1,..., T
101
This representation can be easily modified for panels that do not include a time trend.
(5.2)
219
where i,ts are iid with zero means and finite (heterogeneous) variances.
The null hypothesis is:
H0 :
i = 0 for all i
i.e., yi is an I(1) process for all i
and the alternative is:
H1 :
i < 0, i = 1, 2, , N1,
i = 0, i = N1 + 1, N2 + 2, , N
i.e., i can differ across groups
Both tests allow for heterogeneity of the dynamics and error variances, as in the
panel-based unit root test Levin and Lin (1993) (hereafter LL-test) propose.
By contrast, the LL-test essentially tests the null that each individual series in the
panel is I(I), against the alternative that all the series considered as a panel are
stationary.
It allows for fixed effects and unit-specific time trends in addition to
common time effects, which are an important source of heterogeneity, since the
coefficient on the lagged dependent variable is restricted to be homogeneous across all
units of the panel (Banerjee, 1999). However, Harris and Tzavalis (1999) show that the
LL-test has low power especially when T, the number of time-series observations, is
small.
Similarly, the Monte Carlo experiments IPS (1997) carry out, to investigate the
power performance of the LM-bar and t-bar tests compared with the LL-test, indicate
that the tests proposed by IPS perform much better than the LL-test when the sample
size is small.102 Comparing the LM-bar test and t-bar test, the t-bar test is found to
perform marginally better than the LM-bar test, when T is small and/or N 25. Thus,
102
Breitung (2000) however, finds that both the LL and IPS tests suffer from a loss of power if individual
specific trends are included and, hence, are very sensitive to the specification of the deterministic terms
included in the equations.
220
based on the Monte Carlo results of IPS, the t-bar test is used in this study to test for
unit roots in the panels.
When the disturbances in the underlying DF regressions are not serially

correlated, and are independently and identically distributed across groups, the t-bar
statistic is defined as the average of the N individual DF t-statistics, and is expressed as:
t NT =
1 N
t iT
N i =1
(5.3)
where tiT is the individual t-statistic for testing i = 0. When N is sufficiently large,
i.e., N , IPS (1997) show that the appropriate test statistic converges to a standard
normal distribution under the null hypothesis, and is represented as:
t =
t NT E ( t NT )
Var ( t NT )
N {E NT E (t T )}
Var ( t T )
N ( 0,1)
(5.4)
where E (t NT ) and Var (t NT ) are the common mean and variance across N, respectively.
On the other hand, when the disturbances are serially correlated, the panel unit
root test is based on an Augmented Dickey-Fuller (ADF) regression equation of order
pi :
y i ,t = i + i (t T 2) + i y i ,t 1
pi
+ i , j y i ,t j + i ,t
j =1
, i = 1,..., N ; t = 1,..., T
(5.5)
221
where the number of lagged terms, pi , is chosen to ensure that errors are uncorrelated.103
The t-statistic calculated for this model is:
t NT =
1 N
t iT ( pi , i )
N i =1
(5.6)
When N and T are both finite, the sample distribution of t NT is non-standard and
the exact critical values of the t NT statistic is calculated via stochastic simulation (see
IPS, 1997, Table 4).
Similar to the previous specifications when the errors are serially uncorrelated,
and as N , IPS show that the t-bar statistics converge to standard normal variates,
denoted by: t (p,) where p = (p1, p2, ... , pN) and = (1 , 2 , ... , N). However,
even under the null hypothesis, the sample distributions (and means and variances) of tiT
(pi,i) depend on the nuisance parameter, i , which makes t (p,) not operational.
IPS therefore derive an operational version of t (p,) which does not depend on , as
N and T . This modified standardised version of the t-bar statistic is given by
the equation:
t =
N t NT ( p , )
where t NT ( p , ) =
1
N
1
N
1
N
]}
iN=1 E t iT ( pi ,0) i = 0
Var t iT ( pi ,0) i = 0
N
i =1
(5.7)
N
i =1 t iT ( pi , i ) , t iT ( pi , i ) is the individual t-statistic for testing i =
0, and E[tT(pi,0)] and Var[tT(pi,0)] are the asymptotic values of the mean and variance,
respectively, of the average ADF statistic, computed via stochastic simulations for
various time periods and lags (see IPS, 1997, Table 2). IPS expect t to weakly
converge to standard normal variates under the null hypothesis and diverge under the
103
As mentioned in the previous chapter, ADF test results often depend critically on the number of lagged
differences included in the regression equation. Section 4.4.1 of Chapter 4 details the consequences of
arbitrarily choosing lag lengths and discusses different data-based model selection procedures for
choosing the appropriate lag length. As in the previous chapter, the general to specific (GS) testing
strategy suggested by Campbell and Perron (1991) and Hall (1994) is used to determine pi.
222
alternative hypothesis as both N and T such that N/T k, a finite positive
constant.
In the case where the errors in the different regressions contain a common timespecific component, IPS propose the same t-bar test, but with the individual test
statistics based on cross-sectionally demeaned regressions. That is, assuming that the
disturbance term, it, is composed of two random components:
it = t + vit
(5.8)
where t is a stationary, time-specific common effect, and vit is an idiosyncratic

random effect which is independently distributed across groups (in this case, countries).
To remove the impact of the common effect, the cross-section means should be
subtracted from both sides of the underlying ADF regression (5).
This approach,
however, does not correct for specific dependencies in the errors involving for instance,
pairs of countries within the bigger group.
In this study however, we argue that
demeaning is more appropriate in this context compared to many other applications

because our group of countries, the HPAEs, aside from the fact that they are regionally
connected, have also been hit by common shocks, e.g. the East Asian Crisis that began
in mid-1997. Note too, that at least half of the HPAEs trade takes place within the
region; and while this intra-regional trade has often been blamed for the 1997-1998
crisis, it is now actually fuelling East Asias recovery (The Economist, 2000)
5.5
Panel unit roots based on the IPS (1997) t-bar tests for the more general case
when the errors in the univariate time-series representation are serially correlated
(equations (6) and (7)), for the logarithms of each of the variables original and
demeaned series, with and without trend, are carried out and the results are summarised
in Table 5.2. We opt to use this more general specification of the IPS t-bar test, so as
not to restrict the lagged terms (pi) to be zero for all countries, and instead let pi be
determined based on the general to specific (GS) testing strategy to ensure that errors
223
are uncorrelated. We also consider the demeaned series because we take into account
the more likely possibility that this group of countries has been subjected to common
shocks, and that the impact(s) of such common shocks should be removed by
subtracting the cross-section means from both sides of the underlying ADF regression.
We report and discuss the results based on both the t NT (for finite samples) and
t (standardised t-bar statistics for when N and T tend to infinity) test statistics for
comparison.
The t NT and t test results for the TFP series clearly show that the series,
whether or not it is demeaned and/or includes a time trend, is non-stationary. Only the
t test on the original series with no trend indicates a possibility that TFP is stationary.
For the other series, however, the unit root test results based on the t NT and t test
statistics are mixed, and a conclusive decision on the stationarity of the variables cannot
be made.
The mixed results from the panel unit root tests based on IPS t-bar statistics
( t NT and t ) fail to provide conclusive inference on the order of integration of the
variables of interest. So, while we have hoped that by using panel data, inferences on
the presence of unit roots will be more accurate, our results suggest otherwise. We find
that demeaning the series and whether or not a trend is included in the relevant ADF
regressions could substantially affect the conclusions drawn from the tests. Earlier, we
pointed out why we prefer the demeaned series over the original series; hence, the
remaining issue is whether or not to include a trend term.
224
Table 5.2
Results of IPS Panel Unit Root Tests
t NT
Variable
(in logs)
original
trend
demeaned
no trend
trend
no trend
lnTFP
lnTFP
1.6433
3.4682***
0.4638
2.9332***
2.3843
3.5269***
1.2442
3.4179***
lnTrade
lnTrade
3.118***
5.2309***
1.2331
4.7682***
3.1495***
4.1878***
1.6024
4.2349***
lnDomtax
lnDomtax
2.4418
3.6143***
2.1316*
3.6220***
3.0038**
4.0816***
1.9207
4.1627***
lnNoncomp
lnNoncomp
2.3272
3.3163***
1.8816
3.3787***
2.8581**
3.6006***
1.9804
3.7230***
lnComp
lnComp
2.1361
5.0603***
2.8676**
3.0433***
2.6676*
4.0303***
3.0480***
4.1045***
lnHuman
lnHuman
9.1408***
16.0225***
2.4764**
10.7274***
8.0900***
11.0288***
1.6362
8.6452***
Variable
(in logs)
original
trend
demeaned
no trend
trend
no trend
lnTFP
lnTFP
1.3983
3.4817***
2.6389***
3.5424***
0.6748
3.6323***
0.7021
4.7267***
lnTrade
lnTrade
2.5292**
7.9122***
0.6904
8.0040***
2.6298***
5.2956***
0.2259
6.6994***
lnDomtax
lnDomtax
0.8419
3.8419***
1.6235
5.2197***
2.2647**
5.0916***
1.0365
6.5485***
lnNoncomp
lnNoncomp
0.5322
3.0991***
0.9663
4.6389***
1.9517*
3.8106***
1.2870
5.4306***
lnComp
lnComp
0.0858
7.5494***
3.3547***
3.7825***
1.4605
4.8889***
3.8117***
6.3803***
17.7089***
2.3964**
14.8448***
35.8823***
22.3368***
23.0059***
1%: -3.13 (with trend) ; -2.50, (no trend)
critical values for t NT
5%: -2.82 (with trend); -2.19 (no trend)
10%: -2.67(with trend); -2.04 (no trend)
1% : 2.58; 5% : 1.96; 10%: 1.64
critical values for t
lnHuman
lnHuman
0.4609
17.3121***
***, ** and * reject the null of a unit root at the 1%, 5% and 10% significance levels,
respectively.
225
In our panel unit root tests, we find in favour of a unit root in several series that
do not include a time trend in the univariate time-series representation. This result is
not surprising as Perron (1988) and Elder and Kennedy (2001) note that if the series
actually contains a trend, but this is erroneously omitted, unit root tests tend to be biased
toward finding a unit root. On the other hand, including a trend term, when the series
does not actually have one, reduces the power of the unit root test. For instance,
Gerdtham and Lthgren (2000) note that the main reason for differing results between
unit root tests in a number of empirical studies on health expenditure and GDP is
because some studies base their unit root tests on ADF regressions with a time trend,
while others do not.104 Since the consequences of over-fitting are less harmful than
under-fitting (acknowledging that there is a trade-off between power and size of the
tests), the unit root tests on the series with trend may be preferable. Moreover, trends
are often included to provide a similar test, i.e., so that the distribution of test statistics
is not dependent on any nuisance parameters, in this case, the intercept term.105
Focusing then on the panel unit root test results of the demeaned series with
trend, we find that both t NT and t tests (assuming serially correlated errors) suggest
that TFP is non-stationary, and the rest of the variables: Tradetax, Domtax, Noncomp,
Comp, and Human are all I(0).106 Note, however, that while we have decided that these
results are preferred based on the reasons mentioned earlier, these are still considered
quite fragile, especially since the actual data-generating structure of the variables of
interest is unknown. In such cases, Elder and Kennedy (2001) recommend that unit root
tests simultaneously involve determining what deterministic regressors (intercepts and
trends) should be included. However, this in turn, calls for more tests and/or testing
strategies that are likely to complicate matters more and make unit root test results more
uncertain.
104
Gerdtham and Lthgren (2000) cite for instance that Hansen and King (1996) fail to reject the null of a
unit root for health expenditure and GDP because they include a time trend in their regression equations,
while McCoskey and Selden (1998) reject the null of a unit root because they do not include a time trend.
105
See for instance Patterson (2000), Chapter 6.
106
The
t NT test on the demeaned Comp series, with trend, suggest that the series is I(0), but based on the
t test, it may be I(1).
226
Earlier, we mentioned that our individual country unit root test results would
most likely be different from those obtained in Chapter 4 because of the use of data
converted to international prices (to facilitate pooling of data), and the use of a different
number of time series observations for each country, in order to have a balanced panel.
We report the results of our unit root tests for the individual countries in our panel in
Appendix N.
Comparing these results with those obtained in Chapter 4 (see Table 4.4 and
Appendix I), we find that, the unit root results are very different. From this we are
made aware that even though we are looking at the same countries and the same
variables of interest, the use of data in local or international currencies (particularly for
the TFP variable) and different number of time series observations, considerably affects
the unit root test results. For these reasons, whatever conclusions we make regarding
the order of integration of the variables remain quite fragile.
5.6.
A bounds test approach to the analysis of level relationships
5.6.1
Pesaran et al.s (2001) testing strategy
The results of the panel unit root tests, as well as the individual country unit root
tests, in the previous section and in Chapter 4, show how sensitive unit root test results
can be. There is still much uncertainty as to whether the variables of interest are
actually I(0) or I(1). Given these results, a bounds test approach developed by Pesaran
et al. (2001) for testing level relationships could be useful because it can be applied
irrespective of whether the regressors are I(0) or I(1).
For our purposes, we use Pesaran et al.s bounds tests based on F-type statistics
used to test the significance of the lagged levels of the variables within a univariate
error correction mechanism to determine long-run relations between TFP and our
hypothesised determinants. The F-statistics have non-standard asymptotic distributions
under the null hypothesis that there exists no level relationship, irrespective of whether
the variables of interest are I(0) or I(1), and are analysed against two sets of critical
227
value bounds that cover all possible classifications of the regressors into purely I(0),
purely I(1), or a mixture of I(0)/I(1) variables. If the computed F-statistic falls outside
the critical band, a conclusive decision can be made without needing to know whether
the regressors are I(0) or I(1). That is, if the computed F-statistic falls below the lower
critical band, we fail to reject the null hypothesis of no level relationship, and if the
computed F-statistic falls above the upper critical band, then we reject the null
hypothesis and conclude that there exists a level relationship between our variables of
interest. On the other hand, if the computed F-statistic falls within the bounds, then no
conclusive inference can be made without first knowing the order of integration of the
variables.
Pesaran et al. suggest testing first for the existence of the level relationship
between the regressors using a high-order autoregressive distributed lag (ARDL) model,
to avoid identifying spurious relationships. Once the existence of a level relationship is
confirmed, then the optimal length of the distributed lag is determined based on either
the Akaike (AIC) or Schwarz-Bayesian (SBC) criterion, and the coefficients of the
long-run relations are estimated based on the error correction form of the optimal
ARDL model.
We make some slight modifications to Pesaran et al.s bounds testing strategy,

as applied to our panel. Earlier, we mentioned that one advantage of using panel data
over pure cross-section or time-series estimation, is that, in principle, the use of panel
data will allow the researcher to exploit the dynamics from the cross-sections.
However, because our panel is small, i.e., we have five cross-sections and only 18 timeseries observations for each country, we are unable to exploit more fully the possible
heterogeneity of the underlying economic relationships; instead we pool the data across
the five countries and apply Pesaran et al.s bounds testing strategy.107
107
Our data are arranged in such a way that all the observations for each country are together. It is also
important to specify the lagged variables appropriately in order to keep the lagged variables of a country
in the cross-section separate from the lag values of the next country in the cross-section. An example of a
list of Shazam commands for estimating pooled data with lagged explanatory variables is presented in
Appendix O.
228
By unconditional pooling, we assume that the ith individual outcome at the tth
time period, yit, conditional on variables of interest, say, X , can be viewed as randomly
taken from a common population (Hsiao and Sun, 2000).108
Selecting any set of
economies in an attempt to draw inferences that are valid across countries is often an
arbitrary process. In this study we pool data from five HPAEs for two main reasons.
First, the choice of countries in part reflects the continued attention by researchers to the
economic growth experience of countries in East Asia. From the 1960s to the early
1990s, twenty-three economies in East Asia grew faster than all other regions in the
world, with most of this growth attributable to the performance of the HPAEs. If
growth was randomly distributed, there is roughly one chance in ten thousand that such
a success would be so regionally concentrated (World Bank, 1993a). Second, many
authors have argued that these HPAEs have several shared characteristics aside from
rapid and sustained growth mentioned earlier. Some of the other shared characteristics
of the HPAEs include: very rapid growth of exports and human capital, high levels of
domestic financial savings that sustained high investment levels, judicious government
intervention, and improved productivity. Based on the aforementioned reasons, it is not
entirely implausible to assume a common structure across the HPAEs and rationalize
pooling these economies. Note that due to data constraints mentioned in section 5.3,
only five HPAEs (Indonesia, Malaysia, Singapore, South Korea and Thailand) have
been included in our sample.
We begin our bounds tests by selecting the optimal length of the ARDL model
using a general to specific (GS) testing strategy (discussed in Chapter 4, section 4.5.2)
similar to that suggested by Campbell and Perron (1991) and Hall (1994), starting with
the highest order of lags possible, i.e. of order pmax, where pmax is an upper limit on p,
chosen a priori.
109
Once the optimal length for the distributed lagged difference for
each variable in the model is determined, we then compute the F-statistic and test it
against Pesaran et al.s critical value bounds. If a level relationship is established, we
108
109
Later we will be relaxing this strict assumption using the fixed effects specification in pooling data.
Although we recognise that a pre-testing problem may be inherent in such an approach, we opt to
determine the optimal lag length first, before testing for the existence of the level relationship, mainly
because we do not want to compromise the degrees of freedom available, given the limited number of
time-series observations we have in our pooled data-set.
229
then analyse the coefficients of the long-run relations. We also supplement our analyses
with diagnostic tests.
5.6.2
Bounds tests results
Due to the short span of time-series data available for each country in our panel,
we set the highest order of the lags in our ARDL model at pmax = 5. The error correction
version of our ARDL (5, 5, 5, 5, 5, 5) model in the variables TFP, Tradetax, Domtax,
Noncomp, Comp and Human, where these variables are as previously defined, is given
by:
ln TFPit = 0 + 1 ln TFPit 1 + 2 it 1 + 3 it p + 4 it + u it
(5.9)
p =1
where,
Xit
(lnTradetaxit, lnDomtaxit, lnNoncompit, lnCompit, lnHumanit)'
Zit
(lnTFPit, lnTradetaxit, lnDomtaxit, lnNoncompit, lnCompit,

lnHumanit)'
0 and 1
(lnTFPit , X'it)'
are scalar parameters, assumed to be constant over time and

countries
2,3 and 4 =
are vector parameters, assumed to be constant over time and

countries
denotes the difference operator, and the subscripts i and t denote country i =
(Indonesia, Malaysia, Singapore, South Korea and Thailand), and time t = (1973, 1974,
, 1990) respectively. We assume, at least initially, that the countries in our panel
have common structural characteristics, and that individual country differences have no
significant effect on TFP or, if they do, that these effects are randomly distributed with
zero mean. Ideally, country-specific trend terms should be included in our regression
equation. However, because there are only five cross-sections in the data set and the
number of annual observations for each country is also limited (i.e., t = 18), including
230
country-specific trend terms is not feasible since it would only lead to excessive loss in
degrees of freedom.
Our GS testing strategy to determine the appropriate length of the distributed lag
begins by estimating the ARDL (5, 5, 5, 5, 5, 5) TFP equation (equation (9)). The last
lagged difference term of the regressor with the least significant coefficient is deleted,
and the equation is re-estimated until all the last lagged difference terms of the
regressors are all significant. This results in the choice of an ARDL (0, 0, 3, 0, 3, 0)
specification, and the computed F-statistic for testing the hypothesis that there exists no
level relationship between TFP and the explanatory variables in our equation (2 = 0) is
F(6,52) = 1.0956. This falls outside the lower bound of Pesaran et al.s critical value
bounds, which for the unrestricted intercept and no trend case with k = 6 level
regressors are (2.12, 3.23), (2.45, 3.61) and (3.15, 4.43) at 90%, 95% and 99%
significance levels respectively. This indicates that the hypothesis that there exists no
level relationship in our TFP equation is not rejected. In this case, Pesaran et al. do not
recommend proceeding with the estimation of the error correction form of our ARDL
(0, 0, 3, 0, 3, 0) TFP equation. Nonetheless, for completeness, we also present in
Appendix P the results of estimating the error correction model (ECM) of this ARDL
(0, 0, 3, 0, 3, 0) TFP equation, and Appendix Q presents diagnostic test results for the
preceding estimation.
The estimates presented in Appendix P are consistent with the inference obtained
from the bounds test for the existence of a level relationship. None of the coefficients of
the lagged level regressors is significant. Moreover, diagnostic test results indicate that
while the model is correctly specified, the errors are not normally distributed and a
problem with heteroskedasticity is apparent. Nonetheless, even using the het-consistent
t-ratios to test the significance of the coefficients, only the lagged level
of the Noncomp variable is statistically significant and only at the 10% level. Overall,
these results do not provide sufficient evidence of a long-run relationship between TFP
and our explanatory variables.110
231
5.6.3
Bounds test for the existence of level relationships in a TFP equation with
country-specific dummy variables
In the preceding analysis, we assume that all the countries in our panel data
possess similar structural characteristics and have been subjected to common shocks.
We also assume that structural differences among the HPAEs in our panel do not affect
TFP, or if they do, these effects are randomly distributed with zero mean. If, however,
these differences do matter and are not randomly distributed, then these differences
need to be properly accounted for in our TFP equation. The simplest way of doing so is
by including dummy variables for each of the countries in our panel data set; i.e.,
estimate our pooled data as a fixed effects or dummy variable model (DVM).111 Thus,
our TFP equation with dummy variables is given as:
ln TFPit = 1 ln TFPit 1 + 2 it 1 + 3 it 1 + 4 it
p =1
(5.10)
+ 5 Dit + u it
where the variables and parameters are as previously defined, and the Dits are dummy
variables, such that:
1 for observations for country i (i = 1, . . . , 5)
Dit =
0 otherwise
This means that there is a dummy variable corresponding to each country and
the dummy that corresponds to country i takes the value of one for all observations on
111
There are essentially two ways of estimating coefficients from pooled data sets that would adequately
allow for differences in behaviour over cross-sectional units and/or over time for a given cross-sectional
unit: the fixed effects or dummy variable model (DVM), which treats the disturbance term as fixed and
allows country- and time-specific intercept terms to vary, and the random effects or error components
model, which treats the disturbance term as though it consists of random shocks in the country and time
dimensions. Tests to determine when to use the fixed effects or random effects model based on tests on
the disturbance term have been proposed by Mundlack (1978), Hausman (1978) and Staiger (1988)
among others. However, if the number of cross-sectional units, N, is small as in our pooled data, the
fixed effects model is preferred over the random effects (Judge et al. 1985).
232
country i, but zero for observations on other countries.112
Thus, each country is
characterised by its own particular intercept. However, we still assume homogeneity in

the parameters 1, 2, 3, and 4. Again, as mentioned previously, it would be ideal to
include country-specific time trends, but we are unable to do so to avoid losing valuable
degrees of freedom.
Following the same GS testing strategy to determine the optimal length of the
distributed lag in our ARDL TFP equation, we arrive at an ARDL (1, 5, 3, 3, 2, 1) TFP
equation. We then estimate the ECM form of this equation and present the results in
Table 5.3. Table 5.4 presents diagnostic test results.
The computed F-statistic for the ARDL (1, 5, 3, 3, 2, 1) TFP equation with
dummy variables, F(6,29) = 5.0990, clearly falls outside Pesaran et al.s upper critical
value bounds, which, again for the case with unrestricted intercept and no trend, with
k = 6 level regressors are (2.12, 3.23), (2.45, 3.61) and (3.15, 4.43) at 90%, 95% and
99% significance levels respectively.113
Thus, the null hypothesis of no level
relationship in our TFP equation with dummy variables for each HPAE, is conclusively
rejected, irrespective of whether the variables are all I(0) or all I(1) or a mixture of I(0)
and I(1).
Our estimates also show that four of the six explanatory variables in levels: TFP,
Domtax, Noncomp and Comp, are statistically significant. The lagged levels of TFP and
Noncomp appear to have a significant negative influence on the change in TFP
(lnTFP); while the lagged level of Domtax and Comp appear to have a significant
positive effect. The Tradetax and Human level variables have no significant effect.
112
There are five dummy variables included in equation (10) because it does not include an
additional/separate constant variable.
113
We recognise that for our new equation, where our intercept term is effectively disaggregated into
country-specific dummy variables, the critical values we use to test our hypothesis may be affected; hence
we use Pesaran et al.s critical values with some reservation.
233
Table 5.3

Regressor
Coefficient
Standard error
t-ratio
lnTFPit-1
lnTFPit-1
0.502
0.262
0.130
0.151
3.866***
1.737*
lnTradetaxit-1
lnTradetaxit
lnTradetaxit-1
lnTradetaxit-2
lnTradetaxit-3
lnTradetaxit-4
lnTradetaxit-5
0.000
0.054
0.116
0.055
0.186
0.004
0.099
0.049
0.059
0.080
0.093
0.076
0.0778
0.052
0.001
0.919
1.449
0.591
2.450**
0.058
1.909*
lnDomtaxit-1
lnDomtaxit
lnDomtaxit-1
lnDomtaxit-2
lnDomtaxit-3
0.250
0.128
0.047
0.005
0.274
0.146
0.116
0.116
0.107
0.128
1.710*
1.109
0.410
0.051
2.145**
lnNoncompit-1
lnNoncompit
lnNoncompit-1
lnNoncompit-2
lnNoncompit-3
1.224
0.508
0.652
0.427
0.349
0.302
0.116
0.185
0.124
0.103
4.057***
4.396***
3.525***
3.435***
3.402***
lnCompit-1
lnCompit
lnCompit-1
lnCompit-2
0.540
0.244
0.219
0.156
0.158
0.081
0.105
0.083
3.416***
3.004***
2.081*
1.881*
lnHumanit-1
lnHumanit
lnHumanit-1
0.201
5.112
11.973
0.411
3.843
4.628
0.488
1.330
2.587**
1.959
2.254
2.304
2.489
1.531
2.083
2.169
2.129
2.292
2.068
0.940
1.039
1.083
1.086
0.740
D1
D2
D3
D4
D5
R2 = 0.7531
R2 = 0.4976
F-statistic (6,29) = 5.0990***
Durbin-Watson statistic = 2.1252
Notes:
***, ** and * reject the null hypothesis at 1%, 5% and 10% levels respectively.
234
Table 5.4

Null hypothesis
Test statistics
Jarque-Bera LM test for normality of errors
0.820
(2)
Conclusion
Fail to reject H0
Homoskedasticity
( i )2 on Yi
( i )2 on [Yi ]
( i )2 on ln[Yi ]
( i )2 on X
ln ( i ) on X
2
i on X
7.580***
(1)
1.347
(1)
0.670
(1)
35.001
(30)
30.424
(30)
33.544
(30)
Reject H0
Fail to reject H0
Fail to reject H0
Fail to reject H0
Fail to reject H0
Fail to reject H0
RESET (2)
RESET (3)
RESET (4)
4.417*
(1, 28)
2.437
(2, 27)
1.566
(3, 26)
Reject H0
Fail to reject H0
Fail to reject H0
Notes:
It is uncertain why the level of Domtax would have a positive effect on TFP.
Kneller et al. (1999) and Bleaney et al. (2001) find that non-distortionary taxation (i.e.,
taxes on domestic goods and services, which are assumed not to reduce the returns to
investment) can be growth enhancing if used to finance some productive expenditure.
However, our Domtax measure includes tax components that Kneller et al. and Bleaney
et al. consider to be distortionary (e.g., taxes on income and profit, payroll and
manpower, etc.) that they find, significantly reduces growth.
Of the differenced variables, only those of Noncomp and Comp are found to be
statistically significant. The coefficient of the change in Noncomp points to a negative
235
influence on the change in TFP, while the coefficient of the change in Comp indicates a
positive influence.
Our testing strategy that started with choosing the optimal length of the
distributed lag ensured that at least the last lagged difference term of each variable is
statistically significant. It is interesting to note, however, that while the levels of TFP
and Noncomp indicate a negative influence on the change in TFP, we find that the
lagged changes of these variables generally positively affect the change in TFP. In the
same way, while the levels of Domtax and Comp are positive influences on the change
in TFP, their lagged changes have the opposite (negative) effect. This seems to indicate
that some level variables may initially have a positive (e.g., Domtax and Comp) or
negative (e.g., Noncomp) effect on the change in TFP, but after some time, have a
reverse effect.
For the Tradetax and Human variables, it appears that it takes some time before
their effects on the change in TFP is manifested, i.e. between three to five period lagged
changes for Tradetax, and a one-period lagged change in Human.
In terms of the coefficients on the dummy variables for each of the HPAEs, we
find that none of these is individually statistically significant. However, a joint (F-) test
on the significance of the country-specific dummy variables (H0: 5s are all = 0 against
H1: at least one of the 5s is 0) may be more meaningful than examining the tstatistics associated with the individual coefficients of the dummy variables. This is
because if one follows the procedure of dropping the dummy variables with
insignificant t-statistics, two different parameterisations of the same problem can lead to
different dummy variables being omitted (Judge, 1985 p.521).
The calculated F-
statistic for the joint test on the significance of the dummy variables is F(5,29) = 4.5772,
and this indicates that the coefficients of the dummy variables are significantly different
from zero. This implies that country-specific characteristics do affect our TFP equation.
Such a result may be better confirmed by applying the same bounds test for each
individual country over time. However, because there are not enough time-series data
236
available for the HPAEs, it is not feasible to apply the same bounds test on the
individual HPAEs at this point.114
We find that our regression equation fits reasonably well and also performs well
against a series of diagnostic tests.
Diagnostic tests for normality of error terms,
heteroskedasticity and model mis-specification for the ECM form of the ARDL (1, 5, 3,
3, 2, 1) TFP equation with dummy variables indicate that the errors are normally
distributed and, in general, do not suffer from heteroskedasticity and mis-specification
problems. Overall, we consider the results of our estimation reasonably reliable.
So far, we have only presented the results of the ECM form of the ARDL (1, 5,
3, 3, 2, 1) TFP equation. Earlier we mentioned that the ECM form allows us better
insight on the dynamics involved in our TFP representation. However, from this we can
also obtain the implied long-run level relationship between TFP and the other variables.
Solving our general ECM equation with dummy variables and trend (equation (5.10)),
for lnTFP gives the long-run relationship between TFP and our perceived determinants
as:

ln TFPit = X it + t , where = 2
1
(5.11)
and substituting the results from our ECM estimation, we have (t-statistics are in
parentheses):115
ln TFP = 0.0001 ln Tradetax + 0.4971 ln Domtax 2.4370 ln Noncomp

(0.0006)
(1.7445)
(4.8941)
(5.12)
+ 1.0747 ln Comp + 0.3995 ln Human + intercept i + t

(5.0331)
(0.4855)
114
Recall that in the previous chapter we attempt to analyse long-run relationship among the variables of
interest, but also encounter problems because of lack of data.
115
Since the small sample distribution of the level relationship is unknown, the t-statistics reported
(obtained from non-linear tests) should be interpreted with caution.
237
In analysing the level relationships, we find that the level estimates of Domtax,
Noncomp and Comp are statistically significant. Both Domtax and Comp appear to have
a positive effect on the TFP level, while Noncomp appears to have a highly statistically
significant negative effect. The coefficients of the levels of Tradetax and Human are
not statistically significant.
5.6.4
Bounds test for the existence of level relationships in a TFP equation using detrended variables
In the previous bounds tests, we have not included country-specific trend terms
in our general equation because we are trying to avoid losing valuable degrees of
freedom, given our small panel data set.
However, over the time period being
considered in our analysis, it is unlikely that TFP levels, as well as our other variables,
will not exhibit any trend. There is, therefore, a strong possibility that relationships
estimated from our TFP equations without trend will be at least partly spurious and
influenced by trends in the data. Since we cannot include country-specific time trends
in our equation, one way of addressing this problem is to first de-trend the variables of
interest for each country. It is important to note, however, that in de-trending the
variables we are in effect forcing the variables for each country to have zero means,
hence, we are practically throwing away any long-run information from whatever crosscountry variation there may be. In spite of this reservation, we apply the bounds test
procedure again on equation (5.9), but this time using de-trended variables; so, in effect,
we are including, indirectly, a trend term in equation (5.9).116
We acknowledge that de-trending of the data may yield very different results
compared with the use of trended data. We do not consider that one is more reliable
than the other, or that one is more superior on conceptual grounds. We go through this
116
Note that the trend could arise because it is within the cointegrating space and/or there are quadratic
trends in the data (Patterson, 2000). In this case, the reason for including a trend term in the equation is to
allow for a trend in the long-run level (cointegrating) relationship (Pesaran et al., 2001 and Patterson,
2000).
238
exercise to demonstrate yet another aspect of non-robustness of results where empirical
practice gets little guidance from the theoretical or conceptual literature.
Our GS testing strategy to determine the appropriate length of the distributed lag
of our ARDL (5, 5, 5, 5, 5, 5) TFP equation (equation (5.9)) results in a choice of an
ARDL (5, 0, 3, 0, 3, 3) specification, and the computed F-statistic for testing the
hypothesis that there exists no level relationship between TFP and the explanatory
variables in our equation (2 = 0) is F(6,34) = 5.9973. This clearly falls outside the
upper bound of Pesaran et al.s critical value bounds, which for the unrestricted
intercept and no trend case with k = 6 level regressors are (2.12, 3.23), (2.45, 3.61) and
(3.15, 4.43) at 90%, 95% and 99% significance levels respectively. This indicates that
the hypothesis that there exists no level relationship in our TFP equation is rejected.
The results of estimating the error correction model (ECM) of this ARDL (5, 0, 3, 0, 3,
3) TFP equation and the corresponding diagnostic tests on this equation are presented in
Tables 5.5 and 5.6, respectively.
While we have determined that a level relationship in the ECM form of our
ARDL (5, 0, 3, 0, 3, 3) TFP equation exists, the new set of estimates using de-trended
variables show that only the lagged level terms of TFP and Human are statistically
significant. This is more evident when we draw out the long-run level relationship from
our ECM equation, which is given as (t-statistics are in parentheses):
ln TFP = 0.0160 ln Tradetax + 0.39781 ln Domtax 0.0538 ln Noncomp
(0.0006)
(1.3531)
(0.4066)
(5.13)
0.2907 ln Comp 10.1740 ln Human + intercept i + t

(1.3578)
(2.2473)
239
Table 5.5

(variables are de-trended)
Regressor
Coefficient
Standard error
t-ratio
lnTFPit-1
lnTFPit-1
lnTFPit-2
lnTFPit-3
lnTFPit-4
lnTFPit-5
0.708
0.575
0.373
0.418
0.139
0.278
0.169
0.158
0.167
0.165
0.149
0.142
4.186***
3.633***
2.230**
2.523**
0.935
1.960*
lnTradetaxit-1
lnTradetaxit
0.011
0.057
0.053
0.015
0.213
1.067
lnDomtaxit-1
lnDomtaxit
lnDomtaxit-1
lnDomtaxit-2
lnDomtaxit-3
0.281
0.177
0.139
0.172
0.222
0.209
1.109
0.150
0.147
0.124
1.349
1.625
0.927
1.168
1.801*
lnNoncompit-1
lnNoncompit
0.038
0.172
0.093
0.086
0.410
1.995*
lnCompit-1
lnCompit
lnCompit-1
lnCompit-2
lnCompit-3
0.206
0.014
0.152
0.103
0.163
0.138
0.083
0.091
0.79
0.068
1.488
0.171
1.672*
1.307
2.410**
lnHumanit-1
lnHumanit
lnHumanit-1
lnHumanit-2
lnHumanit-3
7.207
9.282
0.323
2.900
8.841
2.925
4.534
4.282
4.513
4.079
2.464**
2.047**
0.075
0.642
2.168**
Constant
0.006
0.008
0.722
R2 = 0.7259
R2 = 0.5244
F-statistic (6,34) = 5.9973***
Notes:
240
Table 5.6

Null hypothesis
Test statistics
Conclusion
0.1249
(2)
Fail to reject H0
1.351
(1)
0.102
(1)
0.962
(1)
32.931
(25)
22.288
(25)
29.859
(25)
Fail to reject H0
Homoskedasticity
( i )2 on Yi
( i )2 on [Yi ]
( i )2 on ln[Yi ]
( i )2 on X
ln ( i ) on X
2
i on X
Fail to reject H0
Fail to reject H0
Fail to reject H0
Fail to reject H0
Fail to reject H0
RESET (2)
RESET (3)
RESET (4)
7.7491***
(1, 33)
4.5844**
(2, 32)
2.9802**
(3, 31)
Reject H0
Reject H0
Reject H0
Notes:
Our results also suggest that Human has a significant negative effect on TFP.
While our results contradict our hypothesis, this is nonetheless consistent with the
findings of Knight et al. and Islam, that human capital is negatively associated with
growth in their study using panel data. Knight et al. and Islam argue that the time-series
variation in the human capital data is strong enough to outweigh the positive crosssectional effects in the estimation. However, in our estimation, by de-trending the
variables we are effectively throwing out the long-run information from the crosscountry variations and relying on relatively short time-series variation to determine
long-run relationships among our variables of interest; yet, we find that Human appears
to be a negative influence on TFP. Miller and Upadhyay (2000), find that the negative
241
effect of Human on TFP is significant, but only at a 20% significance level. When they
allow for interaction between human capital and openness, they find that Human
positively affects TFP only after a certain level of openness. They propose that too
little openness does not allow a country to reap benefits from its human capital.
Although we have not specifically allowed for any possible interaction between
openness and human capital in our pooled analysis, a negative effect of Human on TFP
may be reasonable, particularly if we note that our openness measure, Tradetax, is
insignificant in our estimation.
Diagnostic tests for normality of error terms, heteroskedasticity and model misspecification for the ECM form of the ARDL (5, 0, 3, 0, 3, 3) TFP equation using detrended variables indicate that the errors are normally distributed and do not suffer from
any heteroskedasticity problem. However, we do find evidence that our TFP equation
is likely mis-specified based on RESET results, hence we remain cautious of any
conclusions we make.
5.7
In this chapter, we attempt to provide further insight as to what factors

essentially drive TFP in five HPAEs using panel data. Recent panel unit root tests
developed by IPS (1997) are applied to determine the order of integration of the data.
However, the results of our panel unit root tests, as well as the individual unit root tests
we perform in this chapter and the previous one, provide no conclusive evidence on the
orders of integration of our variables of interest.
To overcome this situation and enable us to proceed with our analysis, we turn
to Pesaran et al.s (2001) recently developed bounds testing approach to the analysis of
level relationships. The main advantage of this testing strategy is that it can be applied
irrespective of whether the variables of interest are I(0) or I(1), therefore avoiding
difficulties associated with unit root testing (such as those mentioned previously). We
modify Pesaran et al.s testing strategy and apply it to our pooled data.
242
Analysis of the ECM of our ARDL (0, 0, 3, 0, 3, 0) TFP equation without trend
following Pesaran et al.s approach point us to conclude that there exists no long-run
relationship between TFP and any of our perceived determinants. However, if we allow
for country-specific effects in our regression equation (i.e., ARDL (1, 5, 3, 3, 2, 1) TFP
equation), we find strong evidence of a long-run relationship between TFP and our
perceived determinants, irrespective of whether the regressors are I(0) or I(1). Both the
ECM estimation of the underlying ARDL TFP equation and the implied level
relationship indicate that the levels of the two types of government expenditure
variables, Noncomp and Comp, are highly statistically significant influences on TFP.
As hypothesised, the level estimates of Noncomp have a negative effect of TFP levels,
also confirming what we have determined in our time-series analysis (Chapter 4), while
Comp levels have a negative effect.
The level estimates of Domtax, also have a
statistically significant (though to a slightly lesser degree) positive effect on TFP levels.
Completely ignoring time trends in our previous equations, may be

inappropriate, particularly when our variables of interest are likely to exhibit some trend
even over the relatively short time period. However, including country-specific time
trends would lead to loss of valuable degrees of freedom, given our small panel data set.
To overcome this constraint, we opted to de-trend the variables first before estimating
our equations without trend. We acknowledge though that this procedure effectively
throws away any long-run information from cross-country variations. We find that only
our human capital variable has a significant effect on TFP and that, contrary to our
hypothesis, its effect is negative.
Of the three models estimated, we put more weight on the results from
estimating a model that allows for country-specific effects in our TFP equation.
Although the t-statistics on the individual dummy variables are insignificant, the joint
test on the significance of the dummy variables indicates that, as a group, the
coefficients on the dummy variables are significant. This implies that country-specific
effects in our pooled data do matter, and should be accounted for.
243
The information on the long-run relationship between the variables from this
model suggests the importance of these variables in explaining TFP levels in the
HPAEs, particularly the two types of government expenditure. Taken at face value, the
sign and statistical significance of the coefficients on these government expenditure
variables indicate what the effects are of different types of government expenditure on
TFP levels. Expenditures on general public services, defence, and social security and
welfare, which make up our Noncomp variable, adversely affect TFP levels. On the
other hand, expenditures on health, education and economic services, including
expenditures on research, roads, other transportation and communication services,
which make up our Comp variable, positively affect TFP. Although appropriate policy
decisions cannot be made simply on the basis of the statistical significance of these
variables, the information from this study may prove useful as a starting point in
developing policies that could help improve TFP levels in the HPAEs.
244
245
~ Appendix M ~
Complete data set for the estimations (1973-1990)
Country
lnTFPa
lnTadetax
lnDomtax
lnNoncomp
lnComp
lnHuman
Indonesia
7.616089
7.665635
7.689133
7.691919
7.731371
7.789154
7.811782
7.845714
7.947581
7.86891
7.903349
7.911957
7.898438
7.894993
7.877708
7.870745
7.916683
7.979441
-2.50123
-3.00523
-3.03821
-2.91617
-2.96046
-2.86576
-3.00183
-3.30557
-3.59781
-3.75748
-3.94321
-4.0791
-3.88542
-3.45021
-3.16621
-3.70381
-3.65747
-3.54405
-2.22388
-1.97976
-1.91723
-1.84442
-1.84637
-1.8416
-1.71037
-1.60336
-1.64743
-1.70512
-1.77081
-1.81751
-1.74985
-2.07103
-2.03663
-2.00652
-1.97515
-1.79871
-3.1723
-2.71913
-2.54635
-2.45225
-2.46574
-2.32477
-2.18717
-2.17307
-2.26208
-2.3283
-2.35921
-2.44852
-2.31718
-2.35285
-2.63741
-2.61344
-2.60316
-2.49263
-2.74955
-2.67534
-2.38972
-2.28445
-2.42652
-2.42382
-2.34159
-2.10724
-2.18393
-2.23821
-2.22125
-2.33475
-2.3485
-2.70897
-2.68435
-2.62234
-2.67437
-2.64618
4.836988
4.83938
4.841759
4.85178
4.86171
4.871534
4.88127
4.890905
4.89473
4.898541
4.902337
4.906119
4.909886
4.918535
4.927109
4.935603
4.944032
4.952392
Malaysia
8.544091
8.588249
8.468592
8.536409
8.584185
8.605283
8.652028
8.697
8.705601
8.698265
8.686443
8.694476
8.566795
8.462692
8.468136
8.576684
8.672143
8.761962
-2.42974
-2.39418
-2.5207
-2.45256
-2.33491
-2.32422
-2.48446
-2.41454
-2.4875
-2.64907
-2.65042
-2.7085
-2.76067
-2.98477
-3.1537
-3.22934
-3.30498
-3.34215
-2.34698
-2.24182
-1.97591
-2.08677
-1.97378
-1.97597
-2.05057
-1.90771
-1.89897
-1.91821
-1.80487
-1.8633
-1.76568
-1.75404
-2.04713
-2.02392
-2.04637
-1.91205
-2.30918
-2.2873
-2.1076
-2.27433
-2.2043
-2.30354
-2.37897
-2.16845
-1.98625
-1.94601
-2.11645
-2.28465
-2.24116
-2.10872
-2.2618
-2.35085
-2.34293
-2.33313
-2.15226
-1.92663
-1.81376
-1.96991
-1.82342
-1.86405
-2.02871
-1.74726
-1.51254
-1.56351
-1.73359
-1.87955
-1.90949
-1.76851
-2.01647
-2.06247
-2.01858
-1.88237
4.927391
4.934359
4.941278
4.949674
4.957994
4.966251
4.974435
4.982558
4.9895
4.996388
5.003235
5.010028
5.016783
5.02257
5.028325
5.034052
5.039741
5.045397
246
~ Appendix M ~
(continued)
Country
lnTFPa
lnTadetax
lnDomtax
lnNoncomp
lnComp
lnHuman
Singapore
8.983799
8.916663
8.924723
8.911907
8.92186
8.937876
8.961258
8.980036
9.026062
9.019341
9.032944
9.060195
8.95075
8.931555
9.032462
9.162336
9.249989
9.31911
-4.50706
-5.04385
-4.90221
-4.95374
-4.98247
-5.04429
-5.17032
-5.35926
-5.47663
-5.36913
-5.32165
-5.37298
-5.62139
-5.79516
-5.90098
-6.0878
-6.16932
-6.35786
-2.06128
-1.99267
-1.88176
-1.9298
-1.89043
-1.93232
-1.9236
-1.84652
-1.76988
-1.70475
-1.71298
-1.76598
-1.86835
-2.08312
-2.05087
-1.97876
-1.86057
-1.9255
-2.59731
-2.70709
-2.63406
-2.56495
-2.45222
-2.58523
-2.59236
-2.61929
-2.46063
-2.51795
-2.63821
-2.47606
-2.26636
-2.35229
-2.23254
-2.56481
-2.61019
-2.58614
-2.97105
-2.85213
-2.6237
-2.53695
-2.63962
-2.66577
-2.65829
-2.54143
-2.33125
-2.45648
-2.3532
-2.2229
-2.0849
-2.0992
-2.13827
-2.39316
-2.36076
-2.44804
5.004523
5.010442
5.016332
5.017545
5.018762
5.019972
5.021186
5.022393
5.029536
5.036634
5.043677
5.050676
5.057621
5.05448
5.051329
5.048162
5.044992
5.041811
South Korea
8.481496
8.521432
8.520044
8.606201
8.651135
8.698905
8.708712
8.531225
8.537787
8.556447
8.641569
8.694524
8.71839
8.81474
8.909386
9.002918
9.047692
9.108285
-3.57018
-3.5868
-3.3773
-3.14449
-3.05656
-2.85057
-2.93217
-3.17194
-3.30134
-3.24231
-3.01616
-3.09263
-3.0994
-2.85691
-2.7031
-2.7973
-3.00847
-2.9094
-2.37473
-2.25071
-2.13884
-2.06627
-2.11428
-2.11754
-2.07709
-2.05284
-2.03968
-2.02141
-2.03265
-2.08419
-2.08961
-2.11638
-2.11808
-2.04523
-2.02835
-1.98448
-2.79772
-2.69018
-2.5759
-2.53387
-2.52567
-2.47741
-2.55786
-2.43689
-2.42527
-2.3483
-2.51542
-2.60622
-2.59722
-2.62732
-2.71656
-2.74018
-2.68473
-2.69694
-3.0223
-2.89514
-2.78455
-2.75531
-2.93446
-2.98011
-2.68387
-2.83873
-2.86488
-2.75502
-2.82302
-2.82212
-2.80124
-2.88263
-2.87626
-2.88113
-2.7459
-2.72696
5.064119
5.086169
5.107744
5.126087
5.144093
5.161787
5.179168
5.196257
5.205911
5.215479
5.224951
5.234339
5.243634
5.258552
5.273256
5.287742
5.302025
5.316103
247
~ Appendix M ~
(continued)
Country
lnTFPa
lnTadetax
lnDomtax
lnNoncomp
lnComp
lnHuman
Thailand
7.622655
7.583917
7.585092
7.642286
7.704503
7.761881
7.792195
7.773222
7.769605
7.740948
7.807006
7.81375
7.798382
7.809934
7.893708
8.01331
8.109448
8.195731
-2.33589
-2.17073
-2.37209
-2.55542
-2.50528
-2.4878
-2.58199
-2.66372
-2.69303
-2.71335
-2.57323
-2.50264
-2.58078
-2.68907
-2.78305
-2.76938
-2.81948
-2.78282
-2.66947
-2.53427
-2.55704
-2.55256
-2.49348
-2.47781
-2.41547
-2.36499
-2.33551
-2.31705
-2.20815
-2.24027
-2.25111
-2.22597
-2.23928
-2.19428
-2.13559
-2.04297
-3.02127
-3.05543
-3.03042
-2.97223
-2.94665
-2.90053
-2.77756
-2.78145
-2.83761
-2.76136
-2.78094
-2.7752
-2.71359
-2.79221
-2.90679
-3.04805
-3.10542
-3.1114
-2.77824
-2.86914
-2.64644
-2.46607
-2.50064
-2.51899
-2.56082
-2.42838
-2.44018
-2.2916
-2.35174
-2.41728
-2.31985
-2.40137
-2.51008
-2.68632
-2.71863
-2.67551
4.877134
4.877127
4.877112
4.884339
4.891514
4.898645
4.905719
4.912743
4.922634
4.932429
4.942121
4.951727
4.961242
4.967345
4.973411
4.979434
4.985427
4.991385
Notes:
The TFPs calculated in this chapter are slightly different from those calculated in Chapter 3
because the y/l data used in the underlying production function were taken directly from the
PWT data set; while the y/l data used in Chapter 3, although also measured in international
prices, were taken from data provided by Hall and Jones (1999), and only available for the year
1988. See http://www.stanford.edu/~chadj/HallJones400.asc for their complete dataset.
248
~ Appendix N ~
ADF Unit root test results for the individual HPAEs
Country
Indonesia
Malaysia
Variable
(in logs)
Original series
trend
no trend
Demeaned series
trend
no trend
lnTFP
lnTFP
1.9642*
4.3018***
1.5931
4.4056***
0.9816
3.3403***
1.1629
3.0383***
lnTrade
lnTrade
1.7525*
2.7608**
2.1276**
3.4251***
1.4962
3.1849***
2.0655*
3.2239***
lnDomtax
lnDomtax
2.7501**
3.4649***
2.2895**
3.8550***
2.7245**
4.9389***
1.7350*
5.2178***
lnNoncomp
lnNoncomp
3.8228***
0.9584
0.9173
2.6016**
4.4343***
0.9096
1.8184*
2.3720**
lnComp
lnComp
2.4717**
7.9657***
2.7326**
1.4569
5.5973***
1.6711
6.2852***
2.5161**
lnHuman
lnHuman
37.8545***
48.8784***
1.5146
47.1427***
2.0478**
15.4514***
7.2242***
1.1964
lnTFP
lnTFP
2.1107**
2.5799**
2.0750*
2.6035**
2.9616***
2.8026**
0.6773
3.6417***
lnTrade
lnTrade
1.5627
4.1988***
0.8521
3.5414***
1.9616*
4.2697***
1.2216
4.1461***
lnDomtax
lnDomtax
0.9479
4.0277***
2.4501*
3.1449***
4.4184***
3.8359***
2.8348**
4.1120***
lnNoncomp
lnNoncomp
2.2428**
3.6321***
2.2602**
3.6684***
3.2692***
3.8908***
3.6214***
3.9897***
lnComp
lnComp
2.6391**
3.8854***
2.6786**
3.8262***
3.5075***
4.6818***
3.1556***
4.8411***
lnHuman
lnHuman
0.9049
2.8423**
3.1420***
0.6367
2.5787**
2.9798***
0.8873
3.1183***
249
~ Appendix N ~
(continued)
Country
Singapore
South Korea
Variable
(in logs)
Original series
trend
no trend
Demeaned series
trend
no trend
1.6708
2.7194
0.8969
2.5725**
4.3895***
4.3837***
3.2246***
4.0968***
lnTrade
lnTrade
2.4511**
6.8172***
0.5324
6.1396***
1.9408*
5.0461***
0.7821
4.9636***
lnDomtax
lnDomtax
2.4052**
3.2374***
2.4747**
2.9869***
3.4600***
3.5019***
1.9343*
3.7596***
lnNoncomp
lnNoncomp
2.8039**
4.6600***
2.3402**
4.6013***
2.9068***
4.8848***
2.3687**
5.0944***
lnComp
lnComp
1.2404
3.5310***
2.2953**
3.1387***
3.2392***
3.4888***
1.2975
3.5588***
lnHuman
lnHuman
2.0082**
24.8692***
6.4900***
1.1462
30.6428***
33.3419***
1.1490
34.1452***
lnTFP
lnTFP
0.8187
5.4630***
0.4770
2.8660**
2.3045**
3.2970***
1.8794*
2.2806**
lnTrade
lnTrade
5.3616***
8.1175***
3.1731***
6.2881***
6.0344***
4.9304***
1.9877*
5.1947***
lnDomtax
lnDomtax
3.1093***
2.4165**
2.6651**
2.8737**
1.7926*
3.8856***
3.0778***
4.0355***
lnNoncomp
lnNoncomp
2.5348**
3.8876***
1.9055*
3.5668***
4.5334***
3.7837***
0.6557
3.8505***
lnComp
lnComp
3.0435***
5.2191***
4.8810***
3.3076***
1.3359
5.1112***
2.0895**
4.7548***
lnHuman
lnHuman
2.5874**
1.3647
0.2374
1.9174*
2.4592**
1.5143
0.0068
2.1114**
lnTFP
lnTFP
250
~ Appendix N ~
(continued)
Country
Thailand
Variable
(in logs)
Original series
trend
no trend
Demeaned series
trend
no trend
lnTFP
lnTFP
1.6522
2.2818**
0.0249
2.2184**
1.2845
3.8107***
0.7231
4.0323***
lnTrade
lnTrade
4.4261***
4.2601***
1.1843
4.4471***
4.3146***
3.5077***
1.9554*
3.6465***
lnDomtax
lnDomtax
2.9963***
4.9250***
0.7785
5.2498***
2.6238**
4.2458***
0.0214
3.6886***
lnNoncomp
lnNoncomp
0.2353
3.4432***
1.9849*
2.4552**
0.8533
4.5342***
1.4378
3.3083***
lnComp
lnComp
1.2859
4.7004***
1.7506*
3.4872***
2.3300**
5.1986***
2.4121**
4.8517***
lnHuman
lnHuman
2.3490**
2.1568**
0.9979
2.7940***
2.7215**
1.8566*
2.9865***
2.6545**
ADF critical values for T = 18
1%: 2.8784
5%: 2.1009
10%: 1.7341
***, ** and * denote rejection of the null of a unit root at the 1%, 5% and 10% level of
significance, respectively.
251
~ Appendix O ~
A sample list of Shazam commands for estimations of pooled data with
lagged explanatory variables
estimation with a sub-set of time series observations
set the number of time periods
GEN1 NT = 18
generate an index for each cross-section
GENR CSINDEX=SUM(SEAS(NT))
generate a time index for each cross-section
GENR TINDEX=TIME(0) NT*(CSINDEX1)
estimate with lagged variables
estimate over the period 7 to 18
will lose the first observation in each series because of lag and differenced terms
SET NOWARNSKIP
SKIPIF (TINDEX.LT.7)
specify number of lags appropriately, i.e. variable(first.last)
OLS DT DT(1.5) DR(1.5) DD(1.5) DC(1.5) DI(1.5) DH(1.5) LT LR LD LC LI LH DR

DD DC DI DH
STOP
252
~ Appendix P ~
Regressor
Coefficient
Standard error
t-ratio [Het-consistent]
lnTFPit-1
0.007
0.044
0.158 [0.217]
lnTradetaxit-1
lnTradetaxit
0.015
0.090
0.018
0.052
0.791 [1.115]
1.728 [1.824]*
lnDomtaxit-1
lnDomtaxit
lnDomtaxit-1
lnDomtaxit-2
lnDomtaxit-3
0.047
0.144
0.022
0.102
0.164
0.065
0.093
0.100
0.096
0.092
0.729 [0.900]
1.547 [1.772]*
0.217 [0.210]
1.066 [1.101]
1.791* [2.014]*
lnNoncompit-1
lnNoncompit
0.072
0.160
0.058
0.079
1.232 [1.864]*
2.021* [2.095]*
lnCompit-1
lnCompit
lnCompit-1
lnCompit-2
lnCompit-3
0.007
0.031
0.043
0.047
0.125
0.051
0.072
0.070
0.060
0.062
0.133 [0.171]
0.435 [0.354]
0.614 [0.640]
0.790 [0.758]
2.001* [2.134]**
lnHumanit-1
lnHumanit
0.122
4.648
0.176
2.959
0.697 [0.920]
1.571 [1.709]
Constant
0.542
0.547
0.991 [1.321]
R2 = 0.3493
R2 = 0.1365
F-statistic (6,52) = 1.0956
Notes:
Numbers in square brackets are Het-consistent t-ratios. ***, ** and * reject the null hypothesis
at 1%, 5% and 10% levels, respectively,based on Het-consistent t-ratios.
253
~ Appendix Q ~
Null hypothesis
Test statistics
Conclusion
5.0913**
(2)
Reject H0
8.568***
(1)
3.377
(1)
15.555***
(1)
28.086**
(17)
15.093
(17)
27.913**
(17)
Reject H0
Homoskedasticity
( i )2 on Yi
( i )2 on [Yi ]
( i )2 on ln[Yi ]
( i )2 on X
ln ( i ) on X
2
i on X
Fail to reject H0
Reject H0
Reject H0
Fail to reject H0
Reject H0
RESET (2)
RESET (3)
RESET (4)
0.2267
(1, 51)
2.0896
(2, 50)
1.3917
(3, 49)
Fail to reject H0
Fail to reject H0
Fail to reject H0
Notes:
255
~ CHAPTER SIX ~
In this study, we examine total factor productivity (TFP) and its determinants,
with particular emphasis on the experience of the high-performing Asian economies
(HPAEs).
Although the concept of TFP has been extensively analysed in the literature on
economic growth (Griliches, 1995), there is still much uncertainty as to what it really
measures (see Chapter 2). In this study, we do not attempt to come up with a new
interpretation of TFP. Instead, we attempt to empirically identify factors that influence
it, particularly within the experience of the HPAEs.
In Chapter 2, analysis of the more significant growth accounting studies on the

HPAEs shows that the contributions of factor accumulation (FA) and TFP growth in
overall economic growth are quite sensitive to the estimate of the output elasticity of
capital () used, the level of development of the economy, as well as the manner in
which the contribution of human capital to growth is accounted for in the empirical
analysis. We also argue in this chapter that, from a theoretical perspective, any effect of
human capital on overall economic growth is transmitted via its effects on TFP and opt
to exclude human capital from the factors in the underlying production function that we
use to derive our measure of TFP levels. We focus our analysis on levels instead of
growth rates because we are concerned with long-run relationships among our variables
of interest, and levels capture long-run relationships better than growth rates.
We then examine the links between government and openness and the HPAEs
growth experience, as well as the main issues regarding the role of human capital in the
HPAEs growth experience and review the related empirical literature in these areas.
We find that one of the major weaknesses of previous empirical studies is the seeming
indiscriminate use of proxies for government and openness. For instance, the use of
aggregate government expenditure as a proxy for the role of government is not
appropriate because different types of government expenditure may have different
256
effects on the economy. In the next chapters, we make use of different variables to
proxy for government, including a measure of government price controls, a measure of
the extent of government-operated enterprises in the economy, and disaggregated
government expenditure and revenue data.
In Chapter 3, using cross-sectional data, we derive an alternative set of estimates

of TFP levels based on a methodology developed by Hall and Jones (1999). We
examine the correlations between TFP, government, openness and human capital. Our
alternative TFP levels are different from those of Hall and Jones mainly because we
exclude human capital from our underlying production function, and because we use
data in local currencies to avoid any bias in our estimates, which conversion to
international prices may cause.
The multivariate TFP equations estimated in this
chapter use different measures of government intervention and openness, and are
complemented with diagnostic tests to examine the reliability and robustness of our
results.
We conclude in this chapter that differences in TFP levels across countries are
substantially influenced by government intervention. However, contrary to our initial
hypothesis, government intervention in terms of total government consumption
expenditures has a positive influence on TFP.
The other forms of government
intervention, e.g., ownership and control of enterprises and price controls, have, as
hypothesised, negative influences on TFP, but these results are relatively fragile
depending on which other variables are included the regression equation. We find that a
less open economy (measured in terms of either the black market exchange rate
premium or taxes on international trade as a percentage of the total trade sector) is a
significant negative influence on TFP levels. Contrary to our initial hypotheses, we find
that neither human capital nor any interaction between human capital and the degree of
openness, has any significant influence on TFP levels.
In Chapter 4 we continue to examine the links between TFP, government,

openness and human capital using time-series data on five of the HPAEs: Indonesia,
Malaysia, Singapore, South Korea and Thailand.
The focus in this chapter is on
257
establishing the existence of any long-run cointegrating and causal relationships, not
only between TFP and the explanatory variables, but also among the explanatory
variables. For instance, we test what we refer to as the openness-led-growth (OLG)
(instead of export-led-growth) via TFP hypothesis.
That is, we hypothesise that
openness influences economic growth, but essentially through its effect on TFP.
Acknowledging that other factors can also affect the manner in which openness
influences growth, we hypothesise that government intervention and human capital
affect not only TFP but also the degree of openness of an economy. In this chapter, we
use government expenditure disaggregated into non-complementary and complementary
expenditure, as well as domestic tax revenues, as proxies for the different roles of
government. Openness is proxied by taxes on international trade.
The hypothesised cointegrating and causal relationships are analysed within

multivariate cointegration and causality frameworks following Johansens (1988)
cointegration procedure and estimation of error correction models (ECMs), and a
procedure to test for causality, based on Engle and Grangers (1987) two-step
procedure. Again, diagnostic tests, which appear to be lacking in much of the previous
empirical work, are performed to examine the reliability and robustness of our results.
In Chapter 4, we find that government intervention, measured in terms of noncomplementary government expenditure, has a significant negative influence on TFP,
but only for Malaysia.
Human capital, at least for Malaysia, appears to have a
significant positive influence on TFP. There appears to be no strong support for the
openness-led-growth via TFP hypothesis, but government intervention, particularly for
Malaysia and Thailand, is found to influence the degree of openness. Human capital is
also found to have a positive influence on openness in Malaysia. Granger causality tests
for the joint causal effects of the hypothesised determinants, however, show that these
variables jointly Granger cause TFP levels in four of the five HPAEs (i.e., Malaysia,
Singapore, South Korea and Thailand) examined.
In Chapter 5, we attempt to further examine the results from our cross-section

and time-series studies by using panel data and appropriate methods to deal with non-
258
stationarity in the panel (pooled) data.
With panel data, we combine information from
the time-series and cross-sectional dimensions of the data, expecting that inference
about the presence of unit roots will be more accurate when the cross-sectional
dimension of the data is included, particularly when the number of time-series
observations is relatively small. The main contribution of this chapter stems from the
application of one of Pesaran et al.s (2001) bounds testing strategies to test the
existence of long-run level relationships between a dependent variable and a set of
regressors when it is uncertain whether the variables are stationary or non-stationary.
We estimate three TFP models that represent TFP as a function of openness

(measured in terms of taxes on international trade), government (measured in terms of
non-complementary and complementary expenditure), and human capital. The first
model assumes common structural characteristics among the HPAEs in our pooled data
set.
The second allows for structural differences among the HPAEs by including
country-specific dummy variables. The third considers the inclusion of an implicit

trend term in our representation, by de-trending the variables of interest. Of these three
models, we put more weight on the results from estimating a model that allows for
country-specific effects in our TFP equation.
We find evidence of a long-run
relationship between TFP and our hypothesised determinants, regardless of the order of
integration of our variables. However, not all of our hypothesised determinants have a
significant effect on TFP levels. We find that only our government-related variables
have statistically significant influences on TFP. As hypothesised, non-complementary
government expenditure has a statistically significant negative effect, while
complementary expenditure has a statistically significant positive effect. We also find
that, domestic tax revenues have a statistically significant positive effect on TFP levels,
which is contrary to our initial expectations. We suggest that disaggregating domestic
tax revenues into distortionary and non-distortionary components, as in Kneller et al.
(1999) and Bleaney et al. (2001), may provide a clearer insight into this result.
259
This thesis set out to re-examine TFP and its determinants with respect to the
experience of the HPAEs, by testing the robustness of the effects of a selected set of
variables on TFP in a manner consistent with current empirical practice. Of the three
key factors we examine, only the government-related variables appear to have a
significant influence on TFP regardless of the type of data used. However, we also find
that analysis of TFP is highly sensitive to data, choice of proxies for the hypothesised
determinants of TFP, and econometric techniques used.
We acknowledge that the results of this study appear to be somewhat fragile,

suffering persistently from the problem of not having sufficient time-series data in order
to maximize the power of the various econometric techniques that we use to test our
hypotheses.
Nonetheless, all of our estimations are supplemented by a series of
diagnostic tests that we find lacking in many of the previous empirical studies. Hence,
although our results are fragile, it is also possible that results in the existing literature
would also be judged to be fragile if these had reported relevant diagnostics. A larger
panel data set may help improve the precision of our estimates. However, given that
there are only seven HPAEs (excluding Japan), and data for two of the HPAEs (Hong
Kong and Taiwan) are not always available, not much can be done to increase the panel
data for these HPAEs.
Brock and Durlauf (2001, p. 2) note that, the growth literature in economics is
notable for the large gaps that persist between theory and empirics. This is also
particularly true with respect to the literature on TFP. Analysis of TFP remains in its
infancy partly because researchers have yet to reach a consensus regarding its
theoretical meaning. Also, relatively little progress has been made in developing a
theoretical model of the determinants of TFP. However, if we are to advance in our
understanding of the determinants of long-run economic growth beyond factor
accumulation, we have to continue making a concerted effort towards analysing TFP.
The development of a sounder theoretical framework for the empirical analysis of the
determinants of TFP, which other authors have also recently identified as an important
gap, is likely to be an important priority for future work in this area.
261
~ REFERENCES ~
Ahmad, J., Harnhirun, S. and Yang, J. (1997). Exports and economic growth in the
ASEAN countries: Cointegration and causality tests, Rivista Internationale di
Scienze Economiche e Commerciali, 44, 419-430.
Ahn, S. and Hemmings, P. (2000). Policy influences on economic growth in OECD

countries: An evaluation of the evidence, Economic Department Working Paper
No. 246, Organisation for Economic Cooperation and Development.
Akaike, H. (1973). Information theory and an extension of the maximum likelihood

principle, in 2nd International Symposium on Information Theory, Petrov, B.N.
and Csaki, F. (eds), Akademiai Kiado, 267-281.
Alexander, W.R.J., Hansen, P. and Owen, P.D. (1996). Inference on productivity

differentials in multi-sector models of economic growth, Journal of Development
Economics, 51, 315-325.
Amsden, A. (1989). Asias next giant. South Korea and late industrialisation. Oxford
University Press, Oxford and London.
Aoki, M., Kim, H. and Okuno-Fujiwara, M. (1997) (eds). The role of government in
East Asian economic development. Clarendon Press, Oxford.
Arrow, K. J. (1962). The economic implications of learning by doing, Review of

Economic Studies, 29, 155-173
Balassa, B. (1978). Exports and economic growth: further evidence, Journal of

Development Economics, 5, 181-189.
Balassa, B. (1988). The lessons of East Asian development: An overview, Economic

Development and Cultural Change, 36, S273-S290.
262
Balassa, B. (1991). Economic Policies in the Pacific Area Developing Countries.
Macmillan, London.
Baltagi, B. and Griffin, M. (1983). Gasoline demand in the OECD: An application of

pooling and testing procedures, European Economic Review, 22, 117-137.
Baltagi, B. and Kao, C. (2000). Nonstationary panels, cointegration in panels and

dynamic panels: A survey, Texas A&M University and Syracuse University,
mimeo.
Banerjee, A. (1999). Panel data unit roots and cointegration: An overview, Oxford
Bulletin of Economics and Statistics, Special Issue (1999), 607-629.
Banerjee, A., Marcellino, M. and Osbat, C. (2001). Testing for PPP: Should we use
panel methods?, Department of Economics, European University Institute,
mimeo.
Barro, R. (1991). Economic growth in a cross section of countries, Quarterly Journal

of Economics, 106, 407-443.
Barro, R. (1999). Notes on growth accounting, Journal of Economic Growth, 4, 119137.
Barro, R. and Lee, J.W. (1993). International comparisons of educational attainment,

Journal of Monetary Economics, 32, 363-394.
Barro, R. and Lee, J.W. (1994). Sources of economic growth, Carnegie-Rochester

Conference Series on Public Policy, 40, June, 1-46.
Barro, R. and Lee, J.W. (1996).
International measures of schooling years and
schooling quality, American Economic Review, 86, 218-223.
263
Barro, R. and Sala-i-Martin, X. (1995). Economic growth, McGraw-Hill, New York.
Becker, G., Murphy, K. and Tamura, R. (1990). Human capital, fertility and economic
growth, Journal of Political Economy, 98, S13-S37.
Beckman, R.J. and Cook, R.D. (1983). Outliers .s, Technometrics, 25, 119-149.
Behrman, J.R. (1990). Human resource led development?
Review of issues and
evidence, New Delhi and Geneva: ILO-ARTEP.
Belsley, D.A., Kuh, E. and Welsch, R.E. (1980). Regression diagnostics: Identifying
influential data and sources of collinearity. John Wiley and Sons, New York.
Benavot, A. (1989). Education, gender and economic development : A cross-national

study, Sociology of Education, 62, January, 14-32.
Ben-David, D. (1993). Equalising exchange:
Trade liberalisation and income
convergence, Quarterly Journal of Economics, 108, 653-679.
Benhabib, J. and Spiegel, M. (1994).
The role of human capital in economic
development: evidence from aggregate cross-country data, Journal of Monetary

Economics, 34, 143-174.
Benhabib, J. and Spiegel, M. (1997). Cross-country growth regressions, C.V. Starr

Center for Applied Economics, NYU, Research Report 97-20.
Berthlemy, J.C., Dessus, S. and Varoudakis, A. (1997). Human capital and growth:
The role of the trade regime. OECD Development Centre Technical Paper no.
121, Paris.
264
Bhattacharya, A. and Pangestu, M. (1997). Indonesia: Development transformation
and the role of public policy, in Leipziger, D. (ed). Lessons from East Asia,
University of Michigan Press, Michigan.
Bleaney, M. Gemmell, N. and Kneller, R. (2001). Testing the endogenous growth

model: Public expenditures, taxation and growth over the long-run, Canadian
Journal of Economics, 34, 36-57.
Breitung, J. (2000). The local power of some unit root tests for panel data, Advances
in Econometrics, forthcoming.
Breitung, J. and Meyer, W. (1994). Testing for unit roots in panel data: Are wages on
different bargaining levels cointegrated?, Applied Economics, 26, 353-361.
Brock, W.A. and Durlauf, S.N. (2001). Growth empirics and reality, University of
Wisconsin, mimeo.
Campbell, J. and Perron, P. (1991). Pitfalls and opportunities: What macroeconomists

should know about unit roots, in NBER Macroeconomics Annual, 1991,
Blanchard, O.J. and Fischer, S. (eds), MIT Press, Cambridge, MA.
Campos, J.E. and Root, H.L. (1996). The key to the Asian miracle. Making shared
growth credible. The Brookings Institution, Washington, D.C.
Cellini, R. (1997a).
Implications of Solows growth model in the presence of a
stochastic steady state. Journal of Macroeconomics, 19, 135-153.
Cellini, R. (1997b). Growth empirics: evidence from a panel of annual data. Applied
Economics Letters, 4, 347-351.
Chamberlain, G. (1984). Panel data in Griliches, Z. and Intriligator, M. (eds),

Handbook of econometrics, Vol. II, Amsterdam: North-Holland, 1248-1318.
265
Chen, B., Chiang, M. and Kao, C. (1997). International R&D spillovers: An
application of estimation and inference in panel cointegration, Syracuse
University, mimeo.
Chen, E. (1979). Hyper growth in Asian economies. A comparative study of Hong

Kong, Japan, Korea, Singapore and Taiwan. Macmillan Press Ltd., London.
Chow, P.C.Y. (1987). Causality between export growth and industrial performance:
Evidence from the NICs, Journal of Development Economics, 26, 55-63
Cline, W.R. (1987). Can the East Asian model of development be generalised?, in
The Political Economy of Development and Underdevelopment, C.K. Wilber (ed),
Random House Inc., New York.
Coakley, J. and Fuertes, A.M. (1997). New panel unit tests of PPP, Economics
Letters, 57, 17-22.
Coe, D. and Helpman, E. (1995). International R&D spillovers, European Economic

Review, 39, 859-887.
Coe, D., Helpman, E. and Hoffmaister, A. (1997). North-South R&D spillovers,

Economic Journal, 107, 134-149.
Collins, S. and Bosworth, B. (1996). Economic growth in East Asia: Accumulation

versus assimilation, Brookings Paper on Economic Activity, 1996-2, 135-203.
Collins, S., Bosworth, B. and Chen, Y. (1995).
Accounting for differences in
economic growth, Brookings Discussion Papers in International Economics

No. 115.
266
Culver, S.E. and Papell, D.H. (1997). Is there a unit root in the inflation rate?
Evidence from sequential break and panel data models, Journal of Applied
Econometrics, 12, 435-444.
Dawson, J.W. (1998). Institutions, investment and growth: New cross-country and
panel data evidence, Economic Inquiry, 36, 603-619.
de la Fuente, A. (1997). Fiscal policy and growth in the OECD, Centre for Economic
Policy Research Discussion Paper Series, No. 1755.
Deininger, K. and Squire, L. (1997).
Economic growth and income inequality:
reexamining the links, Finance and Development, 38-41.
Demetriades, P., Arestis, O. and Kelly, C. (1998). New evidence on the endogenous
growth debate, University of East London, Department of Economics, Working
Paper Series No. 12.
Demetriades, P.,O. and Hussein, K.A. (1996). Does financial development cause
economic growth?
Time-series evidence from 16 countries, Journal of
Dickey, D.A. and Fuller, W.A. (1979).
Distribution of the estimators for
autoregressive time-series with a unit root Journal of the American Statistical

Association, 74, 427-431.
Dickey, D.A. and Fuller, W.A. (1981). Likelihood ratio statistics for autoregressive
time series with a unit root, Econometrica, 49, 1057-1072.
Dickey, W.A. and Pantula, S.G. (1987). Determining the order of differencing in
autoregressive processes, Journal of Business and Economic Statistics, 5, 455461.
267
Dollar, D. (1990). Outward orientation and growth: An empirical study using a pricebased measure of openness, preliminary draft of a background paper for the
World Development Report 1991, World Bank, Washington, D.C.
Dollar, D. (1992).
rapidly:
Outward-oriented developing economies really do grow more
Evidence from 95 LDCs, 1976-85, Economic Development and
Cultural Change, 40, 523-544.
Doornik, J. A. and Hansen, H. (1994). An omnibus test for univariate and multivariate
normality, Working paper, Nuffield College, Oxford.
Doppelhofer, G., Miller, R. and Sala-i-Martin, X. (2000). Determinants of long-term

growth: A Bayesian averaging of classical estimates (BACE) approach, National
Bureau of Economic Research, Working Paper No. 7750.
Dowling, M. and Summers, P. (1998). Total factor productivity and economic growth
issues for Asia, The Economic Record, 74, 170-185.
Dowrick, S. and Quiggin, J. (1997).
True measures of GDP and convergence,
American Economic Review, 87, 41-64.
Drysdale, P. and Huang, Y. (1997). Technological catch-up and economic growth in

East Asia and the Pacific. Economic Record, 73, 201-211.
Durlauf, S.N. and Quah, D. (1999). The new empirics of economic growth, in
Taylor, J. and Woodford, M. (eds). Handbook of Macroeconomics, Amsterdam,
North Holland.
Easterly, W. and Levine, R. (2000). Its not factor accumulation: Stylised facts and
growth
models,
World
Bank,
mimeo,
available
on
the
(http://www.worldbank.org/research/growth/pdfiles/fact%20final.pdf)
internet
268
Easterly, W. and Rebelo, S. (1993). Fiscal policy and economic growth. Journal of
Monetary Economics, 32, 417-58
Easterly, W., Loayza, N. and Montiel, P. (1997). Has Latin Americas post-reform
growth been disappointing?, Journal of International Economics, 43, 287-311.
Edwards, S. (1992).
Trade orientation, distortions and growth in developing
countries, Journal of Development Economics, 39, 31-57.
Edwards, S.
(1993).
Openness, trade liberalization, and growth in developing
countries, Journal of Economic Literature, 31, 1358-1393.
Edwards, S.
(1997).
Trade policy, growth and income distribution, American
Economic Review, 87, 205-210.
Edwards, S. (1998). Openness, productivity and growth: What do we really know?,

Economic Journal, 108, 383-398.
EIU. (1988). China, Japan and the Asian NICS, Hong Kong & Macau, Singapore,
South Korea, Taiwan: economic structure and analysis. Economist Intelligence
Unit, London.
Elder, J. and Kennedy, P. (2001). Testing for unit roots: What should students be
taught?, Journal of Economic Education, 32, 137-146
Emery, R. (1967). The relation of exports and economic growth, Kyklos, 20, 470-486.
Engelbrecht, H.J. (1998).
Human capital, R&D and North-South knowledge
spillovers, Paper presented to the New Zealand Association of Economists

Annual Conference 1998, 2-4 September 1998, Wellington.
269
Engle, R.F. and Granger, C.W.J. (1987).
Co-integration and error correction:
Representation, estimation and testing, Econometrica, 55, 251-276.
Evans, P. (1997). Government consumption and growth. Economic Inquiry, 35, 209217
Evans, P. (1998). Using panel data to evaluate growth theories, International

Evans, P. and Karras, G. (1996a). Do economies converge? Evidence from a panel of

US states, Review of Economics and Statistics, 78, 384-388.
Evans, P. and Karras, G. (1996b). Are government activities productive? Evidence

from a panel of US states, Review of Economics and Statistics, 76, 1-11.
Feder, G. (1983).
On exports and economic growth, Journal of Development
Economics, 12, 59-73.
Fernandez, C., Ley, E. and Steel, M. (2001). Benchmark priors for Bayesian model
averaging, Journal of Econometrics, 100, 381-427.
Fiebig, D.G. (1992).
Diagnostic checking in practice: the case of outliers and
influential data, University of Sydney, mimeo.
Fishlow, A. Gwin, C., Haggard, S., Rodrik, D. and Wade, R. (1994). Miracle or design?
Lessons from the East Asian experience. Policy Essay No. 11, Overseas
Development Council, Washington, DC.
Fleissig, A.R. and Strauss, J. (1997). Unit root tests on real wage panel data for the
G7, Economics Letters, 56, 149-155.
270
Friedman, M. and Friedman, R. (1980).
Free to choose. A personal statement.
Harcourt Brace Jovanovich, New York and London.
Fuller, W.A. (1976). Introduction to statistical time series, John Wiley and Sons, New
York.
Gemmell, N. (1996).
Evaluating the impacts of human capital stocks and
accumulation on economic growth: Some new evidence, Oxford Bulletin of

Economics and Statistics, 58, 9-28.
Gerdtham, U.-G. and Lthgren, M. (2000). On stationarity and cointegration of

international health expenditure and GDP. Journal of Health Economics, 19,
461-475.
Giles, J.A. and Williams, C.L. (1999). Export-led growth: A survey of the empirical
literature and some noncausality results, Econometrics Working Paper No.
9901, University of Victoria, B.C.
Granger, C.W.J. (1969). Investigating causal relations by econometric models and

cross-spectral methods, Econometrica, 36, 424-438.
Granger, C.W.J. (1988). Recent developments in a concept of causality, Journal of

Econometrics, 39, 199-211.
Greene, W.H. (2000).
Econometric analysis, 4th edition.
Prentice-Hall, Inc. New
Jersey.
Griliches, Z. (1995). The discovery of the residual: An historical note, National

Bureau of Economic Research, Working Paper No. 5348.
Gujarati, D. (1995). Basic econometrics, McGraw-Hill Inc., New York.
271
Gundlach, E. (1997). Openness and economic growth in developing countries,
Weltwirtschaftliches Archiv, 133, 479-496.
Gwartney, J., Lawson, R. and Block, W. (1996). Economic Freedom of the World
1975-1995. Fraser Institute, Vancouver, B.C.
Hall, A. (1994). Testing for a unit root in time series with pretest data-based model
selection, Journal of Business and Economic Statistics, 12, 461-470.
Hall, R. and Jones, C. (1999). Why do some countries produce so much more output
per worker than others?, Quarterly Journal of Economics, 114, 83-116.
Hansen, P. and King, A. (1996). The determinants of health care expenditures: A

cointegration approach, Journal of Health Economics, 15, 127-137.
Harris, R. (1995).
Cointegration analysis in econometric modelling.
Prentice
Hall/Harvester Wheatsheaf, London.
Harris, R.D.F. and Tzavalis, E. (1999). Inference for unit roots in dynamic panels
where the time dimension in fixed, Journal of Econometrics, 91, 201-226.
Harrison, A. (1996). Openness and growth: A time-series, cross-country analysis for

developing countries, Journal of Development Economics, 48, 419-447.
Hartlyn, J. and Morley, S.A. (1986). Political regimes and economic performance in
Latin America, in Hartlyn, J. and Morley, S.A. (eds) Latin American Political
Economy: Financial crisis and political change. Westview Press, London.
Hasan, P. (1976). Korea: Problems and issues in a rapidly growing economy. Johns
Hopkins University Press, Baltimore and London.
272
Hausman, J.A. (1978). Specification tests in econometrics, Econometrica, 46, 12511272.
Helms, L.J. (1985). The effect of state and local taxes on economic growth: A time
series-cross section approach, The Review of Economics and Statistics, 67, 574582.
Hendry, D.F. (1995). Dynamic econometrics, Oxford University Press, Oxford.
Hendry, D.F. and Neale, A.J. (1991). A Monte Carlo study of the effects of structural
breaks on tests for unit roots, in Hackl, P. and Westlung, A. (eds.), Economic
structural change, Springer Verlag, New York, 95-119.
Holden, D. and Perman, R. (1994). Unit roots and cointegration for the economist, in
Rao, B. (ed.), Cointegration for the applied economist. St. Martins Press, New
York, 47-112.
Hsiao, C. (1986). Analysis of panel data. Cambridge University Press, Cambridge.
Hsiao, C. ans Sun, B. (2000). To pool or not to pool panel data, in Krishnakumar, J.
and Ronchetti, E. (eds) Panel data econometrics: furture directions. Papers in
honour of Professor Pietro Balestra. Elsevier, Amsterdam.
Hughes, H. (1995). Why have East Asian countries led economic development, The
Economic Record, 71, 88-104.
Humphries, H. (1999). Investigatng the determinants of economic growth in Taiwan,

M.Com. thesis, University of Otago.
Im, K.S., Pesaran, M.H. and Shin, Y. (1997). Testing for unit roots in heteregeneous
panels. University of Cambridge, mimeo.
273
Islam, I. And Chowdhury, A. (1997). Asia Pacific economies: A survey. Routledge,
London.
Islam, M. (1998). Export expansion and economic growth: testing for cointegration
and causality, Applied Economics, 30, 415-425.
Islam, N. (1995). Growth empirics: A panel data approach, Quarterly Journal of

Economics, 110, November, 1127-1170.
Isuani, E.A. (1994). Social policy and political dynamics in Latin America, in Nagel,
S.S. (ed). Latin American development and public policy. St. Martins Press, New
York.
Johansen, J. (1988). Statistical analysis of cointegration vectors, Journal of Economic

Dynamics and Control, 12, 231-254.
Johansen, J. (1995). A statistical analysis of cointegration for I(2) variables,

Econometric Theory, 11, 25-59.
Johansen, J. and Juselius, K. (1990). Maximum likelihood estimation and inference on

cointegration: With applications to the demand for money, Oxford Bulletin of
Johansen, S. (1992). Determination of cointegration rank in the presence of a linear

trend, Oxford Bulletin of Economics and Statistics, 54, 383-397.
Johnson, C. (1982). MITI and the Japanese miracle. The growth of industrial policy,
1925-1975. Stanford University Press, Stanford, CA.
Joines, D.H. (1981). Estimates of effective marginal tax rates on factor incomes,
Journal of Business 54, 191-226.
274
Jomo K.S. and Gomez, E.T. (1996).
Rents and development in multi-ethnic
Malaysia, in Aoki, M., Kim, H. and Okuno-Fujiwara, M. (eds). The role of

government in East Asian economic development. Clarendon Press, Oxford.
Judge, G.G., Griffiths, W.E., Hill, R.C., Ltkepohl, H. and Lee, T.C. (1985). The
theory and practice of econometrics, John Wiley and Sons, New York.
Jung, W.S. and Marshall, P.J. (1985). Exports, growth and causality in developing
countries, Journal of Development Economics, 18, 1-12.
Kaufmann, D., Kraay, A. and Zoido-Lobatn, P. (1999). Governance matters, World

Bank Policy Research Working Paper No. 2196, World Bank, Washington, D.C.
Kawai, H. (1994). International comparative analysis of economic growth: trade

liberalisation and productivity, The Developing Economies, 32, 373-397.
Kim, J.I. and Lau, L. (1994).
The sources of growth in the East Asian newly
industrialised countries, Journal of the Japanese and International Economies, 8,

235-271.
Kim, J.I. and Lau, L. (1995). The role of human capital in the economic growth of the
East Asian newly industrialised countries, Asia-Pacific Economic Review, 3, 322.
King, R.G. and Levine, R. (1994). Capital fundamentalism, economic development

and economic growth, Carnegie-Rochester Conference Series on Public Policy,
40, 259-292.
Klenow, P. and Rodriguez-Clare, A. (1997). The neo-classical revival in growth

economics: Has it gone too far? in B. Bernanke and J. Rotemberg (eds), 1997
NBER Macroeconomics Annual, MIT Press, Cambridge, MA.
275
Kneller, R., Bleaney, M.F. and Gennell, N. (1999). Fiscal policy and growth: evidence
from OECD countries, Journal of Public Economics, 74, 171-190.
Knight, M., Loayza, N. and Villanueva, D. (1993). Testing the neo-classical theory of
economic growth, IMF Staff Papers, 40, 512-541.
Knowles, S. (2001). Are the Penn World Tables data being misused?, Economics
Letters, 71, 293-298.
Knowles, S. and Owen, P.D. (1997). Education and health in an effective-labour

empirical growth model, Economic Record, 73, 314-328.
Kocherlakota, N. and Yi, K. (1996). A simple time series test of endogenous vs.
exogenous growth models: An application to the United States, Review of
Koester, R. and Kormendi, R. (1989). Taxation, aggregate activity and economic

growth: cross-country evidence on some supply-side hypothesis, Economic
Inquiry, 27, 367-386.
Kormendi, R.C. and Meguire, D.G. (1985). Macroeconomic determinants of growth:

Cross country evidence. Journal of Monetary Economics, 16, 141-163.
Krugman, P. (1994). The myth of Asias miracle, Foreign Affairs, 73, 62-78.
Kurozomi, E. and Yamamoto, T. (2000).
Modified lag augmented vector
autoregressions, Econometric Reviews, 19, 207-231.
Leamer, E. (1983). Lets take the con out of econometrics, American Economic
Review, 73, 31-43.
276
Leamer, E.E. (1985). Vector autoregressions for causal inference, in Brunner, K. and
Meltzer, A. (eds.), Carnegie-Rochester Conference on Public Policy:
Understanding Monetary Regimes, 22, 255-304.
Leamer, E. (1988). Measure of openness, in Baldwin, R. (ed). Trade policy and

empirical analysis, University of Chicago Press, Chicago.
Lee, K., Pesaran, M.H. and Smith, R. (1996). Growth and convergence in a multicountry empirical stochastic Solow model, Journal of Applied Econometrics, 12,
357-392.
Lee, M., Longmire, R., Matyas, L. and Harris, M. (1998). Growth convergence: some
panel data evidence, Applied Economics, 30, 907-912.
Leipziger, D. and Thomas, V. (1993). The lessons of East Asia: An overview of country
experience. World Bank, Washington, D.C
Levin, A. and Lin, C.F. (1992). Unit root tests in panel data: asymptotic and finite
sample properties. University of California, San Diego, mimeo.
Levin, A. and Lin, C.F. (1993). Unit root tests in panel data: new results, Discussion
paper 56, University of California, San Diego.
Levine, R. and Renelt, D. (1992). A sensitivity analysis of cross-country growth

regressions, American Economic Review, 82, 942-963.
Lipsey, R.G. and Carlaw, K. (2001). What does total factor productivity measure?,
Simon Fraser University and University of Canterbury, mimeo.
Loayza, N.V. (1994). A test of the international convergence hypothesis using panel
data. Policy Research Working Paper No. 1333, World Bank.
277
Lorgelly, P.K. and Owen, P.D. (1999). The effect of female and male schooling in
economic growth in the Barro-Lee model, Empirical Economics, 24, 537-557.
Lucas, R.E. Jr. (1988). On the mechanics of economic development, Journal of

Monetary Economics, 22, 3-42.
Lynde, C. and Richmond, J. (1993).
Public capital and long-run costs in U.K.
manufacturing, Economic Journal, 103, 880-893.
MacDonald, R. (1996). Panel unit root tests and real exchange rates, Economics
Letters, 50, 7-11.
Macintyre, A. (1994).
Business, government and development: Northeast and
Southeast Asian comparisons, in Macintyre, A. (ed) Business and government in

industrialising Asia. Allen and Unwin Pty. Ltd., St. Leonards, NSW.
Maddala, G.S. and Kim, I-M. (1998). Unit roots, cointegration and structural change,
Cambridge University Press, Cambridge.
Maddison, A. (1987).
Growth and slowdown in advanced capitalist economies:
Techniques of quantitative assessment, Journal of Economic Literature, 25,

649-698.
Maddison, A. (1995). Explaining the economic performance of nations: Essays in time

and space. Edward Elgar Publishing Co., Vermont.
Mairesse, J. (1990). Time-series and cross-sectional estimates on panel data: Why are
they different and why should they be equal? in Hartog, J, Ridder, G. and
Theeuwes, J. (eds), Panel data and labour market studies, Amsterdam: NorthHolland, 81-95.
278
Mankiw, G., Romer, D. and Weil, D. (1992). A contribution to the empirics of
economic growth, Quarterly Journal of Economics, 107, 407-437.
McGhee, J.W. (1985). Introductory statistics, West Publishing Company, St. Paul,
Minnesota.
McKinnon, J.G. (1991). Critical values for cointegration tests, Chapter 13 in Engle,
R.F. and Granger, C.W.J. (eds.). Long-run economic relationships: Readings in
cointegration, Oxford University Press, Oxford.
McKoskey, S. and Kao, C. (1998). How does your output grow? A panel data
investigation of the relationship between urbanisation and growth, United States
Naval Academy and Syracuse University, mimeo.
McKoskey, S. and Selden, T.M. (1998). Health care expenditures and GDP: Panel
unit root test results, Journal of Health Economics, 17, 369-376.
Mendoza, E., Milesi-Ferretti, G. and Asea, P. (1997). On the ineffectiveness of tax

policy in altering long-run growth: Harbergers superneutrality conjecture,
Journal of Public Economics, 66, 99-126
Mendoza, E., Razin, A. and Tesar, L. (1994). Effective tax rates in macroeconomics.
Cross-country estimates of tax rates on factor incomes and consumption,
Journal of Monetary Economics, 34, 297-323.
Miller, S.M. (1996). A note on cross-country growth regressions, Applied Economics,

28, 1019-1026.
Miller, S.M. and Upadhyay, M.P. (2000). The effects of openness, trade orientation,
and human capital on total factor productivity, Journal of Development
Economics, 63, 399-423.
279
Mincer, J.(1974). Schooling, experience and earnings. Columbia University Press, New
York.
Mishra, V.
(1995).
Indonesia: Adjustment in the 1980s,
in Shaw, T.M. (ed).
Economic restructuring in East Asia and India. Perspectives on policy reform,

Macmillan Press Ltd., London.
Mundlack, Y. (1978).
On the pooling of time-series and cross-section data,
Economterica, 46, 69-85.
Nehru, V. and Dhareshwar, A. (1994). New estimates of total factor productivity

growth for developing and industrial countries, World Bank Policy Research
Working paper No. 1313, World Bank, Washington, D.C.
Nehru, V. Swanson, E. and Dubey, A. (1995). A new database on human capital stock
in developing and industrial countries: Sources, methodology and results,
Journal of Development Economics, 46, 379-401.
Nelson, R.R. and Phelps, E.S. (1966). Investments in humans, technological diffusion,
and economic growth, American Economic Review, 56, 69-75.
Ng, D. (2000). Granger causality in possibly integrated vector autoregressive models:

A Monte Carlo comparison and an empirical application to government
expenditure and economic growth.
M.Com.(Hons.) thesis, University of
Canterbury.
Oh, K-Y. (1996). Purchasing power parity and unit root tests using panel data,
Journal of International Money and Finance, 15, 405-418.
Osterwald-Lenum, M. (1992). A note on quantiles of the asymptotic distribution of

the maximum likelihood cointegration rank test statistic, Oxford Bulletin of
280
Oxley, L. (1993). Cointegration, causality and export-led growth in Portugal, 18651985, Economics Letters, 43, 163-166.
Pack, H. and Page, J.M. Jr. (1994). Accumulation, exports and growth in the high
performing Asian economies, Carnegie Rochester Conference Series on Public
Policy, 40, 199-236.
Page, J.M. (1994). The East Asian miracle: An introduction, World Development, 22
(4), 615-625
Pantula, S.G. (1989). Testing for unit roots in time series data, Econometric Theory,
5, 256-271.
Patterson, K. (2000). An introduction to applied econometrics. A time-series approach.

St. Martins Press, New York.
Pedroni, P. (1996). Human capital, endogenous growth and cointegration for multicountry panels, Indiana University, mimeo.
Pedroni, P. (1997). Panel cointegration: Asymptotic and finite sample properties of

pooled time-series tests with an application to the PPP hypothesis, Indiana
University, mimeo.
Pernia, E.M. (1993). Economic growth performance of Indonesia, the Philippines and
Thailand:
The human resource dimension, in Ogawa et al. (eds), Human
resources in development along the Asia-Pacific rim. Oxford University Press,

Singapore.
Perron, P. (1988). Trends and random walks in macroeconomic time series, Journal
of Economic Dynamics and Control, 12, 297-332.
281
Perron, P. (1989). The great crash, the oil price shock and the unit root hypothesis,
Econometrica, 57, 1361-1401.
Pesaran, M.H. and Pesaran, B. (1997). Microfit 4.0 windows version, Oxford University
Press, Oxford.
Pesaran, M.H., Shin, Y. and Smith, R.J. (2001). Bounds testing approaches to the
analysis of level relationships, Journal of Applied Econometrics, 16, 289-326.
Pesaran, M.H., and Smith, R.J. (1995).
Estimating long-run relationships from
dynamic heterogeneous panel, Journal of Econometrics, 68, 79-113.
Phillips, P.C.B. (1987). Time series regression with a unit root, Econometrica, 55,
277-301.
Phillips, P.C.B. and Moon, H.R. (1999). Nonstationary panel data analysis: An
overview of some recent developments, Yale University, Cowles Foundation for
Research in Economics Discussion Paper No. 1221.
Podivinsky, J.M. (1990). Testing misspecified cointegrating relationships, Working

Paper No. 9, Monash University, Australia.
Prescott, E.C. (1998). Needed: A theory of total factor productivity, International

Pritchett, L. (1996). Measuring outward orientation in developing countries: Can it be

done?, Journal of Development Economics, 49, 307-335.
Pritchett, L. (1997). Where has all the education gone?, World Bank Policy Research
Working Paper No. 1581, World Bank, Washington, D.C.
282
Psacharopoulos, G. (1994). Returns to investment in education: A global update,
World Development, 22, 1325-1343.
Quah, D. (1994). Exploiting cross section variation for unit root inference in dynamic
data, Economic Letters, 44, 9-19.
Ram, R. (1986). Government size and economic growth: a new framework and some
evidence from cross-section and time-series data, American Economic Review,
76, 191-203.
Reichlin, L. (1989).
Structural change and unit root econometrics, Economics
Letters, 31, 231-233.
Reimers, H.E. (1992). Comparisons of tests for multivariate cointegration, Statistical

Papers, 33, 335-359.
Robertson, P.E. (1999). On growth accounting, growth theory and the East Asian
miracle, School of Economics Working Paper no. 98/12, University of New
South Wales.
Rodriguez, F. and Rodrik, D. (2000). Trade policy and economic growth: A skeptics
guide to the cross-national evidence, forthcoming in NBER Macroeconomics
Annual.
Romer, P. (1986). Increasing returns and long-run growth, Journal of Political

Economy, 94, 1002-1037.
Romer, P. (1990). Human capital and growth: theory and evidence, CarnegieRocehster Conference Series on Public Policy, 32, 251-286.
Root, H.L. (1996). Small countries, big lessons: governance and the rise of East Asia,
Oxford University Press, Hong Kong.
283
Sachs, J. and Warner, A. (1995).
Economic reform and the process of global
integration, Brookings Papers on Economic Activity, 1, 1-118.
Said, S.E. and Dickey, D.A. (1984). Testing for unit roots in autoregressive-moving
average models of unkown order, Biometrika, 71, 599-608.
Sala-i-Martin, X. (1997). I just ran two million regressions, American Economic

Review, Papers and Proceedings, 87, 178-83.
Sarel, M. (1997). Growth and productivity in ASEAN countries, IMF Working Paper
97/97, International Monetary Fund, Washington.
Schwarz, G. (1978). Estimating the dimension of a model, The annals of statistics, 6,

461-464.
Seater, J. (1985). On the construction of marginal federal and social security tax rates
in the U.S., Journal of Monetary Economics, 15, 121-135.
Seguino, S. (2000).
Gender inequality and economic growth: A cross-country
analysis, World Development, 28, 1211-1230.
Sen , A. and Srivastava, M. (1990).
Regression analysis.
Theory, methods and
applications. Springer-Verlag, New York.
Senhadji, A. (1999). Sources of economic growth: An extensive growth accounting

exercise, IMF Working Paper No. 99/77, International Monetary Fund
Serletis, A. (1992). Export growth and Canadian economic development, Journal of

284
Sims, C. (1972). Money, income and causality, American Economic Review, 62, 540552.
Singh, A. (1993). Asian economic success and Latin American failure in the 1980s:
new analyses and future policy implications, International Review of Applied
Economics, 7, 267-289.
Snodgrass, D.R. (1998).
Education in Korea and Malaysia, in Rowen, H. (ed).
Behind East Asian growth. The political and social foundations of prosperity,
Routledge, London and New York.
Solow, R. (1956). A contribution to the theory of economic growth, Quarterly

Journal of Economics, 70, 65-94.
Staiger, R.W. (1988). A specification test of the Hecksher-Ohlin Theory, Journal of

International Economics, 25, 129-141.
Stiglitz, J.E. (1996).
Some lessons from the East Asian miracle, World Bank
Research Observer, 11, 151-177.
Summers, R. and Heston, A. (1988). A new set of international comparisons of real

product and price estimates for 130 countries, 1950-85, Review of Income and
Wealth, 34, 1-26.
Summers, R. and Heston, A. (1991). The Penn World Table (Mark 5): An Expanded
Set of International Comparisons 1950-1988, Quarterly Journal of Economics,
106, 327-368.
Swan, T.W. (1956). Economic growth and capital accumulation. Economic Record,
66, 334-361.
285
Tallman, E. and Wang, P. (1994). Human capital and endogenous growth: Evidence
from Taiwan, Journal of Monetary Economics, 34, 101-124.
Tan, G. (1992). The newly industrialising countries of Asia. Times Academic Press,
Singapore.
Taylor, A.M. (1995). Growth and convergence in the Asia-Pacific Region: On the role
of openness, trade and migration.
National Bureau of Economic Research,
Working Paper No. 5276.
Temple, J. (1997). St. Adam and the dragons: Neo-classical economics and the East
Asian miracle, Oxford Development Studies, 25, 279-300.
Temple, J. (1999a). The new growth evidence, Journal of Economic Literature, 37,
112-156.
Temple, J. (1999b). A positive effect of human capital on growth, Economics Letters,

65, 131-134.
Temple, J.R.W. (2001). Generalizations that arent?
Evidence on education and
growth, Department of Economics, University of Bristol, mimeo.

The Economist. (2000). Happy neighbours, August 26th 2000, p.71.
Thomas, V. and Wang, Y. (1993). Government policy and productivity growth: Is

East Asia an exception?. The Lessons of East Asia, World Bank, Washington,
D.C.
Thomas, V. and Wang, Y. (1996).
Distortions, Interventions, and Productivity
Growth: Is East Asia Different?, Economic Development and Cultural Change,

44, 265-288.
286
Tipton, F.B. (1998). The rise of Asia. Economics, society and politics in contemporary
Asia. Macmillan Press Ltd., London.
Toda, H.Y. (1995).
Finite sample performance of likelihood ratio tests for
cointegration ranks in vector autoregression, Econometric Theory, 11, 10151032.
Toda, H.Y. and Phillips, P.C.B. (1993).
Vector autoregressions and causality,
Econometrica, 61, 1367-1393.
Toda, H.Y. and Phillips, P.C.B. (1994).
Vector autoregressions and causality: A
theoretical overview and simulation study, Econometric Reviews, 13, 259-285.
Toda, H.Y. and Yamamoto. T. (1995). Statistical inference in vector autoregressions

with possibly integrated processes, Journal of Econometrics, 66, 225-250.
Uzawa, H. (1965). Optimum technical change in an aggregative model of economic

growth, International Economic Review, 6, 18-31.
Villegas, B. M. (1998). A speedy recovery for the Philippine economy, University of

Asia and the Pacific, mimeo.
Wade, R. (1990). Governing the market. Economic theory and the role of government
in East Asian industrialisation.
Princeton University Press, Princeton, New
Jersey.
Warr, P.G. and Nidhiprabha, B. (1996). Thailands macroeconomic miracle: Stable

adjustment and sustained growth. Oxford University Press, Kuala Lumpur.
White, H. (1980). A heteroskedasticity-consistent covariance matrix estimator and a

direct test for heteroskedasticity, Econometrica, 48, 817-838.
287
White, K.J. (1993). SHAZAM users reference manual version 7.0, McGraw-Hill Book
Company, New York.
Wiarda, H.J. (1987). Latin America at the crossroads: Debt, development and the
future. Westview Press Inc., Colorado.
World Bank. (1988). World development report, 1988, Washington D.C.
World Bank. (1991).
World development report 1991. Oxford University Press,
Oxford.
World Bank. (1993a). The East Asian miracle: Economic growth and public policy,
Oxford University Press, Washington, DC.
World Bank. (1993b). Sustaining rapid development in East Asia and the Pacific.
Development in Practice Series, World Bank, Washington, DC.
Young, A. (1992). A tale of two cities: Factor accumulation and technical change in
Hong Kong and Singapore, NBER Macroeconomics Annual, 7, 13-54.
Young, A. (1994). Lessons from the East Asian NICs: A contrarian view. European
Young, A. (1995). The tyranny of numbers: Confronting the statistical realities of the
East Asian growth experience, Quarterly Journal of Economics, 110, 641-680.
288

PHD Thesis

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

PHD Thesis

Hochgeladen von

Copyright:

Verfügbare Formate

The Determinants of Total Factor Productivity:

The High-Performing Asian Economies Revisited

Arlene Locsin Garces Ozanne

A thesis submitted for the degree of

We hypothesise that the degree of openness of an economy, the different roles

Results from our cross-sectional analysis show that government intervention,

There is some evidence that the degree of openness of the economy

Our time-series analysis of the individual HPAEs,

Granger causality tests for the joint causal effects of the

I am also grateful for the financial support from the Macandrew-Stout

Any errors and omissions in this thesis are my sole responsibility.

2. The HPAEs Economic Growth Experience

2.2. What is TFP?

2.2.1. FA and TFP growth

2.2.2. TFP and its determinants within a single growth equation

2.2.3. TFP from growth accounting

2.2.4. TFP as residuals from a production function

2.3. Government and Economic Growth

2.3.1. To intervene or not to intervene

2.3.2. Government intervention in the HPAEs

2.3.3. Analysing the empirical evidence

2.4. Openness and Economic Growth

2.4.1. Establishing the link between openness and economic growth

2.4.2. Analysing the empirical evidence

2.5. Human capital and growth in the HPAEs

3.2. Estimating TFP levels

3.2.1. Hall and Jones (1999) TFP levels

3.2.2. An alternative measure of TFP levels

3.3. Correlation analyses

3.3.2. Correlation analysis results

3.4. TFP and its determinants

3.4.2. Regression results

3.5. Robustness and Diagnostic tests

3.5.1. Robustness tests

3.5.2. Diagnostic tests

3.5.3. TFP estimations without the influential observations

3.6. Summary and conclusions

Explanation of GOE Ratings

Explanation of Price Ratings

1988 TFP levels (in logs)

TFP estimation results, using TFP from a production function

Diagnostic test results

TFP estimation results, using TFP from a production function

Diagnostic test results

Diagnostic test results

Diagnostic test results

List of countries included in the main TFP regressions

4. Testing for Cointegration and Causality for the HPAEs

4.3. TFP levels for the HPAEs

4.4. The relationship between the Granger causality concept

4.4.1. The Granger causality concept

4.4.2. Cointegration, error correction models (ECM) and causation

4.5. Analysis of the time-series properties

4.5.1. Unit root testing procedure

4.5.2. Choosing the appropriate lag length

4.5.3. Unit root test results

4.6. Cointegration and error correction model (ECM) representation

4.6.1. Johansens multivariate cointegration and ECM representation

4.6.2. Testing for cointegration in the HPAEs

4.6.3. OLS estimation results for South Korea

4.7. Testing for Granger causality

4.8. Summary and conclusions

H.4 South Korea